blob: 33939e72bc372a0d4b312d444c5f2bad7eede9f2 [file] [log] [blame]
Christoph Lameter81819f02007-05-06 14:49:36 -07001/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
Christoph Lameter881db7f2011-06-01 12:25:53 -05005 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
Christoph Lameter81819f02007-05-06 14:49:36 -07007 *
Christoph Lametercde53532008-07-04 09:59:22 -07008 * (C) 2007 SGI, Christoph Lameter
Christoph Lameter881db7f2011-06-01 12:25:53 -05009 * (C) 2011 Linux Foundation, Christoph Lameter
Christoph Lameter81819f02007-05-06 14:49:36 -070010 */
11
12#include <linux/mm.h>
Nick Piggin1eb5ac62009-05-05 19:13:44 +100013#include <linux/swap.h> /* struct reclaim_state */
Christoph Lameter81819f02007-05-06 14:49:36 -070014#include <linux/module.h>
15#include <linux/bit_spinlock.h>
16#include <linux/interrupt.h>
17#include <linux/bitops.h>
18#include <linux/slab.h>
Christoph Lameter97d06602012-07-06 15:25:11 -050019#include "slab.h"
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +040020#include <linux/proc_fs.h>
Andrew Morton3ac38fa2013-04-29 15:08:06 -070021#include <linux/notifier.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070022#include <linux/seq_file.h>
Vegard Nossum5a896d92008-04-04 00:54:48 +020023#include <linux/kmemcheck.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024#include <linux/cpu.h>
25#include <linux/cpuset.h>
26#include <linux/mempolicy.h>
27#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070028#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070029#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070030#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070031#include <linux/math64.h>
Akinobu Mita773ff602008-12-23 19:37:01 +090032#include <linux/fault-inject.h>
Pekka Enbergbfa71452011-07-07 22:47:01 +030033#include <linux/stacktrace.h>
Christoph Lameter4de900b2012-01-30 15:53:51 -060034#include <linux/prefetch.h>
Glauber Costa2633d7a2012-12-18 14:22:34 -080035#include <linux/memcontrol.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070036
Richard Kennedy4a923792010-10-21 10:29:19 +010037#include <trace/events/kmem.h>
38
Mel Gorman072bb0a2012-07-31 16:43:58 -070039#include "internal.h"
40
Christoph Lameter81819f02007-05-06 14:49:36 -070041/*
42 * Lock order:
Christoph Lameter18004c52012-07-06 15:25:12 -050043 * 1. slab_mutex (Global Mutex)
Christoph Lameter881db7f2011-06-01 12:25:53 -050044 * 2. node->list_lock
45 * 3. slab_lock(page) (Only on some arches and for debugging)
Christoph Lameter81819f02007-05-06 14:49:36 -070046 *
Christoph Lameter18004c52012-07-06 15:25:12 -050047 * slab_mutex
Christoph Lameter881db7f2011-06-01 12:25:53 -050048 *
Christoph Lameter18004c52012-07-06 15:25:12 -050049 * The role of the slab_mutex is to protect the list of all the slabs
Christoph Lameter881db7f2011-06-01 12:25:53 -050050 * and to synchronize major metadata changes to slab cache structures.
51 *
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
58 *
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -070064 *
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
70 *
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
75 * the list lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070076 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
80 *
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
83 *
Christoph Lameter672bba32007-05-09 02:32:39 -070084 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070086 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070087 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070089 *
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
93 *
94 * Overloading of page flags that are otherwise used for LRU management.
95 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070096 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
104 *
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700108 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -0700109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -0700111 *
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700114 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700115 */
116
Christoph Lameteraf537b02010-07-09 14:07:14 -0500117static inline int kmem_cache_debug(struct kmem_cache *s)
118{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700119#ifdef CONFIG_SLUB_DEBUG
Christoph Lameteraf537b02010-07-09 14:07:14 -0500120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
Christoph Lameter5577bd82007-05-16 22:10:56 -0700121#else
Christoph Lameteraf537b02010-07-09 14:07:14 -0500122 return 0;
Christoph Lameter5577bd82007-05-16 22:10:56 -0700123#endif
Christoph Lameteraf537b02010-07-09 14:07:14 -0500124}
Christoph Lameter5577bd82007-05-16 22:10:56 -0700125
Joonsoo Kim345c9052013-06-19 14:05:52 +0900126static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127{
128#ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
130#else
131 return false;
132#endif
133}
134
Christoph Lameter81819f02007-05-06 14:49:36 -0700135/*
136 * Issues still to be resolved:
137 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700140 * - Variable sizing of the per node arrays
141 */
142
143/* Enable to test recovery from slab corruption on boot */
144#undef SLUB_RESILIENCY_TEST
145
Christoph Lameterb789ef52011-06-01 12:25:49 -0500146/* Enable to log cmpxchg failures */
147#undef SLUB_DEBUG_CMPXCHG
148
Christoph Lameter81819f02007-05-06 14:49:36 -0700149/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
152 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800153#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700154
Christoph Lameter2086d262007-05-06 14:49:46 -0700155/*
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
Zhi Yong Wu721ae222013-11-08 20:47:37 +0800158 * sort the partial list by the number of objects in use.
Christoph Lameter2086d262007-05-06 14:49:46 -0700159 */
160#define MAX_PARTIAL 10
161
Christoph Lameter81819f02007-05-06 14:49:36 -0700162#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700164
Christoph Lameter81819f02007-05-06 14:49:36 -0700165/*
David Rientjes3de47212009-07-27 18:30:35 -0700166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
168 * metadata.
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700169 */
David Rientjes3de47212009-07-27 18:30:35 -0700170#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700171
172/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700173 * Set of flags that will prevent slab merging
174 */
175#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +0300176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
177 SLAB_FAILSLAB)
Christoph Lameter81819f02007-05-06 14:49:36 -0700178
179#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
Vegard Nossum5a896d92008-04-04 00:54:48 +0200180 SLAB_CACHE_DMA | SLAB_NOTRACK)
Christoph Lameter81819f02007-05-06 14:49:36 -0700181
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400182#define OO_SHIFT 16
183#define OO_MASK ((1 << OO_SHIFT) - 1)
Christoph Lameter50d5c412011-06-01 12:25:45 -0500184#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400185
Christoph Lameter81819f02007-05-06 14:49:36 -0700186/* Internal SLUB flags */
Christoph Lameterf90ec392010-07-09 14:07:11 -0500187#define __OBJECT_POISON 0x80000000UL /* Poison object */
Christoph Lameterb789ef52011-06-01 12:25:49 -0500188#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
Christoph Lameter81819f02007-05-06 14:49:36 -0700189
Christoph Lameter81819f02007-05-06 14:49:36 -0700190#ifdef CONFIG_SMP
191static struct notifier_block slab_notifier;
192#endif
193
Christoph Lameter02cbc872007-05-09 02:32:43 -0700194/*
195 * Tracking user of a slab.
196 */
Ben Greeard6543e32011-07-07 11:36:36 -0700197#define TRACK_ADDRS_COUNT 16
Christoph Lameter02cbc872007-05-09 02:32:43 -0700198struct track {
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300199 unsigned long addr; /* Called from address */
Ben Greeard6543e32011-07-07 11:36:36 -0700200#ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202#endif
Christoph Lameter02cbc872007-05-09 02:32:43 -0700203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
206};
207
208enum track_item { TRACK_ALLOC, TRACK_FREE };
209
Christoph Lameterab4d5ed2010-10-05 13:57:26 -0500210#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -0700211static int sysfs_slab_add(struct kmem_cache *);
212static int sysfs_slab_alias(struct kmem_cache *, const char *);
213static void sysfs_slab_remove(struct kmem_cache *);
Glauber Costa107dab52012-12-18 14:23:05 -0800214static void memcg_propagate_slab_attrs(struct kmem_cache *s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700215#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700216static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
217static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
218 { return 0; }
Christoph Lameterdb265ec2012-09-04 23:18:33 +0000219static inline void sysfs_slab_remove(struct kmem_cache *s) { }
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800220
Glauber Costa107dab52012-12-18 14:23:05 -0800221static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
Christoph Lameter81819f02007-05-06 14:49:36 -0700222#endif
223
Christoph Lameter4fdccdf2011-03-22 13:35:00 -0500224static inline void stat(const struct kmem_cache *s, enum stat_item si)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800225{
226#ifdef CONFIG_SLUB_STATS
Christoph Lameter84e554e62009-12-18 16:26:23 -0600227 __this_cpu_inc(s->cpu_slab->stat[si]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800228#endif
229}
230
Christoph Lameter81819f02007-05-06 14:49:36 -0700231/********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
234
Christoph Lameter81819f02007-05-06 14:49:36 -0700235static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
236{
Christoph Lameter81819f02007-05-06 14:49:36 -0700237 return s->node[node];
Christoph Lameter81819f02007-05-06 14:49:36 -0700238}
239
Christoph Lameter6446faa2008-02-15 23:45:26 -0800240/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700241static inline int check_valid_pointer(struct kmem_cache *s,
242 struct page *page, const void *object)
243{
244 void *base;
245
Christoph Lametera973e9d2008-03-01 13:40:44 -0800246 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700247 return 1;
248
Christoph Lametera973e9d2008-03-01 13:40:44 -0800249 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300250 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700251 (object - base) % s->size) {
252 return 0;
253 }
254
255 return 1;
256}
257
Christoph Lameter7656c722007-05-09 02:32:40 -0700258static inline void *get_freepointer(struct kmem_cache *s, void *object)
259{
260 return *(void **)(object + s->offset);
261}
262
Eric Dumazet0ad95002011-12-16 16:25:34 +0100263static void prefetch_freepointer(const struct kmem_cache *s, void *object)
264{
265 prefetch(object + s->offset);
266}
267
Christoph Lameter1393d9a2011-05-16 15:26:08 -0500268static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
269{
270 void *p;
271
272#ifdef CONFIG_DEBUG_PAGEALLOC
273 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
274#else
275 p = get_freepointer(s, object);
276#endif
277 return p;
278}
279
Christoph Lameter7656c722007-05-09 02:32:40 -0700280static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
281{
282 *(void **)(object + s->offset) = fp;
283}
284
285/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300286#define for_each_object(__p, __s, __addr, __objects) \
287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700288 __p += (__s)->size)
289
Christoph Lameter7656c722007-05-09 02:32:40 -0700290/* Determine object index from a given position */
291static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292{
293 return (p - addr) / s->size;
294}
295
Mariusz Kozlowskid71f6062011-02-26 20:10:26 +0100296static inline size_t slab_ksize(const struct kmem_cache *s)
297{
298#ifdef CONFIG_SLUB_DEBUG
299 /*
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
302 */
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500304 return s->object_size;
Mariusz Kozlowskid71f6062011-02-26 20:10:26 +0100305
306#endif
307 /*
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
311 */
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
313 return s->inuse;
314 /*
315 * Else we can use all the padding etc for the allocation
316 */
317 return s->size;
318}
319
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800320static inline int order_objects(int order, unsigned long size, int reserved)
321{
322 return ((PAGE_SIZE << order) - reserved) / size;
323}
324
Christoph Lameter834f3d12008-04-14 19:11:31 +0300325static inline struct kmem_cache_order_objects oo_make(int order,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800326 unsigned long size, int reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300327{
328 struct kmem_cache_order_objects x = {
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800329 (order << OO_SHIFT) + order_objects(order, size, reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300330 };
331
332 return x;
333}
334
335static inline int oo_order(struct kmem_cache_order_objects x)
336{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400337 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300338}
339
340static inline int oo_objects(struct kmem_cache_order_objects x)
341{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400342 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300343}
344
Christoph Lameter881db7f2011-06-01 12:25:53 -0500345/*
346 * Per slab locking using the pagelock
347 */
348static __always_inline void slab_lock(struct page *page)
349{
350 bit_spin_lock(PG_locked, &page->flags);
351}
352
353static __always_inline void slab_unlock(struct page *page)
354{
355 __bit_spin_unlock(PG_locked, &page->flags);
356}
357
Dave Hansena0320862014-01-30 15:46:09 -0800358static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
359{
360 struct page tmp;
361 tmp.counters = counters_new;
362 /*
363 * page->counters can cover frozen/inuse/objects as well
364 * as page->_count. If we assign to ->counters directly
365 * we run the risk of losing updates to page->_count, so
366 * be careful and only assign to the fields we need.
367 */
368 page->frozen = tmp.frozen;
369 page->inuse = tmp.inuse;
370 page->objects = tmp.objects;
371}
372
Christoph Lameter1d071712011-07-14 12:49:12 -0500373/* Interrupts must be disabled (for the fallback code to work right) */
374static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500375 void *freelist_old, unsigned long counters_old,
376 void *freelist_new, unsigned long counters_new,
377 const char *n)
378{
Christoph Lameter1d071712011-07-14 12:49:12 -0500379 VM_BUG_ON(!irqs_disabled());
Heiko Carstens25654092012-01-12 17:17:33 -0800380#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
381 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -0500382 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000383 if (cmpxchg_double(&page->freelist, &page->counters,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500384 freelist_old, counters_old,
385 freelist_new, counters_new))
386 return 1;
387 } else
388#endif
389 {
Christoph Lameter881db7f2011-06-01 12:25:53 -0500390 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800391 if (page->freelist == freelist_old &&
392 page->counters == counters_old) {
Christoph Lameterb789ef52011-06-01 12:25:49 -0500393 page->freelist = freelist_new;
Dave Hansena0320862014-01-30 15:46:09 -0800394 set_page_slub_counters(page, counters_new);
Christoph Lameter881db7f2011-06-01 12:25:53 -0500395 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500396 return 1;
397 }
Christoph Lameter881db7f2011-06-01 12:25:53 -0500398 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500399 }
400
401 cpu_relax();
402 stat(s, CMPXCHG_DOUBLE_FAIL);
403
404#ifdef SLUB_DEBUG_CMPXCHG
405 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
406#endif
407
408 return 0;
409}
410
Christoph Lameter1d071712011-07-14 12:49:12 -0500411static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
412 void *freelist_old, unsigned long counters_old,
413 void *freelist_new, unsigned long counters_new,
414 const char *n)
415{
Heiko Carstens25654092012-01-12 17:17:33 -0800416#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
417 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameter1d071712011-07-14 12:49:12 -0500418 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000419 if (cmpxchg_double(&page->freelist, &page->counters,
Christoph Lameter1d071712011-07-14 12:49:12 -0500420 freelist_old, counters_old,
421 freelist_new, counters_new))
422 return 1;
423 } else
424#endif
425 {
426 unsigned long flags;
427
428 local_irq_save(flags);
429 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800430 if (page->freelist == freelist_old &&
431 page->counters == counters_old) {
Christoph Lameter1d071712011-07-14 12:49:12 -0500432 page->freelist = freelist_new;
Dave Hansena0320862014-01-30 15:46:09 -0800433 set_page_slub_counters(page, counters_new);
Christoph Lameter1d071712011-07-14 12:49:12 -0500434 slab_unlock(page);
435 local_irq_restore(flags);
436 return 1;
437 }
438 slab_unlock(page);
439 local_irq_restore(flags);
440 }
441
442 cpu_relax();
443 stat(s, CMPXCHG_DOUBLE_FAIL);
444
445#ifdef SLUB_DEBUG_CMPXCHG
446 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
447#endif
448
449 return 0;
450}
451
Christoph Lameter41ecc552007-05-09 02:32:44 -0700452#ifdef CONFIG_SLUB_DEBUG
453/*
Christoph Lameter5f80b132011-04-15 14:48:13 -0500454 * Determine a map of object in use on a page.
455 *
Christoph Lameter881db7f2011-06-01 12:25:53 -0500456 * Node listlock must be held to guarantee that the page does
Christoph Lameter5f80b132011-04-15 14:48:13 -0500457 * not vanish from under us.
458 */
459static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
460{
461 void *p;
462 void *addr = page_address(page);
463
464 for (p = page->freelist; p; p = get_freepointer(s, p))
465 set_bit(slab_index(p, s, addr), map);
466}
467
Christoph Lameter41ecc552007-05-09 02:32:44 -0700468/*
469 * Debug settings:
470 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700471#ifdef CONFIG_SLUB_DEBUG_ON
472static int slub_debug = DEBUG_DEFAULT_FLAGS;
473#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700474static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700475#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700476
477static char *slub_debug_slabs;
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700478static int disable_higher_order_debug;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700479
Christoph Lameter7656c722007-05-09 02:32:40 -0700480/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700481 * Object debugging
482 */
483static void print_section(char *text, u8 *addr, unsigned int length)
484{
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200485 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
486 length, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700487}
488
Christoph Lameter81819f02007-05-06 14:49:36 -0700489static struct track *get_track(struct kmem_cache *s, void *object,
490 enum track_item alloc)
491{
492 struct track *p;
493
494 if (s->offset)
495 p = object + s->offset + sizeof(void *);
496 else
497 p = object + s->inuse;
498
499 return p + alloc;
500}
501
502static void set_track(struct kmem_cache *s, void *object,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300503 enum track_item alloc, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700504{
Akinobu Mita1a00df42009-03-07 00:36:21 +0900505 struct track *p = get_track(s, object, alloc);
Christoph Lameter81819f02007-05-06 14:49:36 -0700506
Christoph Lameter81819f02007-05-06 14:49:36 -0700507 if (addr) {
Ben Greeard6543e32011-07-07 11:36:36 -0700508#ifdef CONFIG_STACKTRACE
509 struct stack_trace trace;
510 int i;
511
512 trace.nr_entries = 0;
513 trace.max_entries = TRACK_ADDRS_COUNT;
514 trace.entries = p->addrs;
515 trace.skip = 3;
516 save_stack_trace(&trace);
517
518 /* See rant in lockdep.c */
519 if (trace.nr_entries != 0 &&
520 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
521 trace.nr_entries--;
522
523 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
524 p->addrs[i] = 0;
525#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700526 p->addr = addr;
527 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400528 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700529 p->when = jiffies;
530 } else
531 memset(p, 0, sizeof(struct track));
532}
533
Christoph Lameter81819f02007-05-06 14:49:36 -0700534static void init_tracking(struct kmem_cache *s, void *object)
535{
Christoph Lameter24922682007-07-17 04:03:18 -0700536 if (!(s->flags & SLAB_STORE_USER))
537 return;
538
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300539 set_track(s, object, TRACK_FREE, 0UL);
540 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700541}
542
543static void print_track(const char *s, struct track *t)
544{
545 if (!t->addr)
546 return;
547
Linus Torvalds7daf7052008-07-14 12:12:53 -0700548 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300549 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Ben Greeard6543e32011-07-07 11:36:36 -0700550#ifdef CONFIG_STACKTRACE
551 {
552 int i;
553 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
554 if (t->addrs[i])
555 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
556 else
557 break;
558 }
559#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700560}
561
Christoph Lameter24922682007-07-17 04:03:18 -0700562static void print_tracking(struct kmem_cache *s, void *object)
563{
564 if (!(s->flags & SLAB_STORE_USER))
565 return;
566
567 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
568 print_track("Freed", get_track(s, object, TRACK_FREE));
569}
570
571static void print_page_info(struct page *page)
572{
Chen Gangd0e0ac92013-07-15 09:05:29 +0800573 printk(KERN_ERR
574 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
575 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700576
577}
578
579static void slab_bug(struct kmem_cache *s, char *fmt, ...)
580{
581 va_list args;
582 char buf[100];
583
584 va_start(args, fmt);
585 vsnprintf(buf, sizeof(buf), fmt, args);
586 va_end(args);
587 printk(KERN_ERR "========================================"
588 "=====================================\n");
Dave Jones265d47e2011-11-15 15:04:00 -0800589 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700590 printk(KERN_ERR "----------------------------------------"
591 "-------------------------------------\n\n");
Dave Jones645df232012-09-18 15:54:12 -0400592
Rusty Russell373d4d02013-01-21 17:17:39 +1030593 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Christoph Lameter24922682007-07-17 04:03:18 -0700594}
595
596static void slab_fix(struct kmem_cache *s, char *fmt, ...)
597{
598 va_list args;
599 char buf[100];
600
601 va_start(args, fmt);
602 vsnprintf(buf, sizeof(buf), fmt, args);
603 va_end(args);
604 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
605}
606
607static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700608{
609 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800610 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700611
612 print_tracking(s, p);
613
614 print_page_info(page);
615
616 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
617 p, p - addr, get_freepointer(s, p));
618
619 if (p > addr + 16)
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200620 print_section("Bytes b4 ", p - 16, 16);
Christoph Lameter24922682007-07-17 04:03:18 -0700621
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500622 print_section("Object ", p, min_t(unsigned long, s->object_size,
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200623 PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700624 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500625 print_section("Redzone ", p + s->object_size,
626 s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700627
Christoph Lameter81819f02007-05-06 14:49:36 -0700628 if (s->offset)
629 off = s->offset + sizeof(void *);
630 else
631 off = s->inuse;
632
Christoph Lameter24922682007-07-17 04:03:18 -0700633 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700634 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700635
636 if (off != s->size)
637 /* Beginning of the filler is the free pointer */
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200638 print_section("Padding ", p + off, s->size - off);
Christoph Lameter24922682007-07-17 04:03:18 -0700639
640 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700641}
642
643static void object_err(struct kmem_cache *s, struct page *page,
644 u8 *object, char *reason)
645{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700646 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700647 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700648}
649
Chen Gangd0e0ac92013-07-15 09:05:29 +0800650static void slab_err(struct kmem_cache *s, struct page *page,
651 const char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700652{
653 va_list args;
654 char buf[100];
655
Christoph Lameter24922682007-07-17 04:03:18 -0700656 va_start(args, fmt);
657 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700658 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700659 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700660 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700661 dump_stack();
662}
663
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500664static void init_object(struct kmem_cache *s, void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700665{
666 u8 *p = object;
667
668 if (s->flags & __OBJECT_POISON) {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500669 memset(p, POISON_FREE, s->object_size - 1);
670 p[s->object_size - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700671 }
672
673 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500674 memset(p + s->object_size, val, s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700675}
676
Christoph Lameter24922682007-07-17 04:03:18 -0700677static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
678 void *from, void *to)
679{
680 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
681 memset(from, data, to - from);
682}
683
684static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
685 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800686 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700687{
688 u8 *fault;
689 u8 *end;
690
Akinobu Mita798248202011-10-31 17:08:07 -0700691 fault = memchr_inv(start, value, bytes);
Christoph Lameter24922682007-07-17 04:03:18 -0700692 if (!fault)
693 return 1;
694
695 end = start + bytes;
696 while (end > fault && end[-1] == value)
697 end--;
698
699 slab_bug(s, "%s overwritten", what);
700 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
701 fault, end - 1, fault[0], value);
702 print_trailer(s, page, object);
703
704 restore_bytes(s, what, value, fault, end);
705 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700706}
707
Christoph Lameter81819f02007-05-06 14:49:36 -0700708/*
709 * Object layout:
710 *
711 * object address
712 * Bytes of the object to be managed.
713 * If the freepointer may overlay the object then the free
714 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700715 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700716 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
717 * 0xa5 (POISON_END)
718 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500719 * object + s->object_size
Christoph Lameter81819f02007-05-06 14:49:36 -0700720 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700721 * Padding is extended by another word if Redzoning is enabled and
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500722 * object_size == inuse.
Christoph Lameter672bba32007-05-09 02:32:39 -0700723 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700724 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
725 * 0xcc (RED_ACTIVE) for objects in use.
726 *
727 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700728 * Meta data starts here.
729 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700730 * A. Free pointer (if we cannot overwrite object on free)
731 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700732 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800733 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700734 * before the word boundary.
735 *
736 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700737 *
738 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700739 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700740 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500741 * If slabcaches are merged then the object_size and inuse boundaries are mostly
Christoph Lameter672bba32007-05-09 02:32:39 -0700742 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700743 * may be used with merged slabcaches.
744 */
745
Christoph Lameter81819f02007-05-06 14:49:36 -0700746static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
747{
748 unsigned long off = s->inuse; /* The end of info */
749
750 if (s->offset)
751 /* Freepointer is placed after the object. */
752 off += sizeof(void *);
753
754 if (s->flags & SLAB_STORE_USER)
755 /* We also have user information there */
756 off += 2 * sizeof(struct track);
757
758 if (s->size == off)
759 return 1;
760
Christoph Lameter24922682007-07-17 04:03:18 -0700761 return check_bytes_and_report(s, page, p, "Object padding",
762 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700763}
764
Christoph Lameter39b26462008-04-14 19:11:30 +0300765/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700766static int slab_pad_check(struct kmem_cache *s, struct page *page)
767{
Christoph Lameter24922682007-07-17 04:03:18 -0700768 u8 *start;
769 u8 *fault;
770 u8 *end;
771 int length;
772 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700773
774 if (!(s->flags & SLAB_POISON))
775 return 1;
776
Christoph Lametera973e9d2008-03-01 13:40:44 -0800777 start = page_address(page);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800778 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
Christoph Lameter39b26462008-04-14 19:11:30 +0300779 end = start + length;
780 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700781 if (!remainder)
782 return 1;
783
Akinobu Mita798248202011-10-31 17:08:07 -0700784 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700785 if (!fault)
786 return 1;
787 while (end > fault && end[-1] == POISON_INUSE)
788 end--;
789
790 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200791 print_section("Padding ", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700792
Eric Dumazet8a3d2712009-09-03 16:08:06 +0200793 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
Christoph Lameter24922682007-07-17 04:03:18 -0700794 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700795}
796
797static int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500798 void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700799{
800 u8 *p = object;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500801 u8 *endobject = object + s->object_size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700802
803 if (s->flags & SLAB_RED_ZONE) {
Christoph Lameter24922682007-07-17 04:03:18 -0700804 if (!check_bytes_and_report(s, page, object, "Redzone",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500805 endobject, val, s->inuse - s->object_size))
Christoph Lameter81819f02007-05-06 14:49:36 -0700806 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700807 } else {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500808 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800809 check_bytes_and_report(s, page, p, "Alignment padding",
Chen Gangd0e0ac92013-07-15 09:05:29 +0800810 endobject, POISON_INUSE,
811 s->inuse - s->object_size);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800812 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700813 }
814
815 if (s->flags & SLAB_POISON) {
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500816 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700817 (!check_bytes_and_report(s, page, p, "Poison", p,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500818 POISON_FREE, s->object_size - 1) ||
Christoph Lameter24922682007-07-17 04:03:18 -0700819 !check_bytes_and_report(s, page, p, "Poison",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500820 p + s->object_size - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700821 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700822 /*
823 * check_pad_bytes cleans up on its own.
824 */
825 check_pad_bytes(s, page, p);
826 }
827
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500828 if (!s->offset && val == SLUB_RED_ACTIVE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700829 /*
830 * Object and freepointer overlap. Cannot check
831 * freepointer while object is allocated.
832 */
833 return 1;
834
835 /* Check free pointer validity */
836 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
837 object_err(s, page, p, "Freepointer corrupt");
838 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100839 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700840 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700841 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700842 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800843 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700844 return 0;
845 }
846 return 1;
847}
848
849static int check_slab(struct kmem_cache *s, struct page *page)
850{
Christoph Lameter39b26462008-04-14 19:11:30 +0300851 int maxobj;
852
Christoph Lameter81819f02007-05-06 14:49:36 -0700853 VM_BUG_ON(!irqs_disabled());
854
855 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700856 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700857 return 0;
858 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300859
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800860 maxobj = order_objects(compound_order(page), s->size, s->reserved);
Christoph Lameter39b26462008-04-14 19:11:30 +0300861 if (page->objects > maxobj) {
862 slab_err(s, page, "objects %u > max %u",
863 s->name, page->objects, maxobj);
864 return 0;
865 }
866 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700867 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300868 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700869 return 0;
870 }
871 /* Slab_pad_check fixes things up after itself */
872 slab_pad_check(s, page);
873 return 1;
874}
875
876/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700877 * Determine if a certain object on a page is on the freelist. Must hold the
878 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700879 */
880static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
881{
882 int nr = 0;
Christoph Lameter881db7f2011-06-01 12:25:53 -0500883 void *fp;
Christoph Lameter81819f02007-05-06 14:49:36 -0700884 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300885 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700886
Christoph Lameter881db7f2011-06-01 12:25:53 -0500887 fp = page->freelist;
Christoph Lameter39b26462008-04-14 19:11:30 +0300888 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700889 if (fp == search)
890 return 1;
891 if (!check_valid_pointer(s, page, fp)) {
892 if (object) {
893 object_err(s, page, object,
894 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800895 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700896 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700897 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800898 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300899 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700900 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700901 return 0;
902 }
903 break;
904 }
905 object = fp;
906 fp = get_freepointer(s, object);
907 nr++;
908 }
909
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800910 max_objects = order_objects(compound_order(page), s->size, s->reserved);
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400911 if (max_objects > MAX_OBJS_PER_PAGE)
912 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300913
914 if (page->objects != max_objects) {
915 slab_err(s, page, "Wrong number of objects. Found %d but "
916 "should be %d", page->objects, max_objects);
917 page->objects = max_objects;
918 slab_fix(s, "Number of objects adjusted.");
919 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300920 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700921 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300922 "counted were %d", page->inuse, page->objects - nr);
923 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700924 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700925 }
926 return search == NULL;
927}
928
Christoph Lameter0121c6192008-04-29 16:11:12 -0700929static void trace(struct kmem_cache *s, struct page *page, void *object,
930 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700931{
932 if (s->flags & SLAB_TRACE) {
933 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
934 s->name,
935 alloc ? "alloc" : "free",
936 object, page->inuse,
937 page->freelist);
938
939 if (!alloc)
Chen Gangd0e0ac92013-07-15 09:05:29 +0800940 print_section("Object ", (void *)object,
941 s->object_size);
Christoph Lameter3ec09742007-05-16 22:11:00 -0700942
943 dump_stack();
944 }
945}
946
Christoph Lameter643b1132007-05-06 14:49:42 -0700947/*
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500948 * Hooks for other subsystems that check memory allocations. In a typical
949 * production configuration these hooks all should produce no code at all.
950 */
Roman Bobnievd56791b2013-10-08 15:58:57 -0700951static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
952{
953 kmemleak_alloc(ptr, size, 1, flags);
954}
955
956static inline void kfree_hook(const void *x)
957{
958 kmemleak_free(x);
959}
960
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500961static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
962{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500963 flags &= gfp_allowed_mask;
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500964 lockdep_trace_alloc(flags);
965 might_sleep_if(flags & __GFP_WAIT);
966
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500967 return should_failslab(s->object_size, flags, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500968}
969
Chen Gangd0e0ac92013-07-15 09:05:29 +0800970static inline void slab_post_alloc_hook(struct kmem_cache *s,
971 gfp_t flags, void *object)
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500972{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500973 flags &= gfp_allowed_mask;
Eric Dumazetb3d41882011-02-14 18:35:22 +0100974 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500975 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500976}
977
978static inline void slab_free_hook(struct kmem_cache *s, void *x)
979{
980 kmemleak_free_recursive(x, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500981
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600982 /*
Xie XiuQid1756172013-10-18 09:12:43 +0800983 * Trouble is that we may no longer disable interrupts in the fast path
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600984 * So in order to make the debug calls that expect irqs to be
985 * disabled we need to disable interrupts temporarily.
986 */
987#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
988 {
989 unsigned long flags;
990
991 local_irq_save(flags);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500992 kmemcheck_slab_free(s, x, s->object_size);
993 debug_check_no_locks_freed(x, s->object_size);
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600994 local_irq_restore(flags);
995 }
996#endif
Thomas Gleixnerf9b615d2011-03-24 21:26:46 +0200997 if (!(s->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500998 debug_check_no_obj_freed(x, s->object_size);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500999}
1000
1001/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001002 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -07001003 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001004static void add_full(struct kmem_cache *s,
1005 struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -07001006{
Christoph Lameter643b1132007-05-06 14:49:42 -07001007 if (!(s->flags & SLAB_STORE_USER))
1008 return;
1009
David Rientjes255d0882014-02-10 14:25:39 -08001010 lockdep_assert_held(&n->list_lock);
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001011 list_add(&page->lru, &n->full);
1012}
Christoph Lameter643b1132007-05-06 14:49:42 -07001013
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001014static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001015{
1016 if (!(s->flags & SLAB_STORE_USER))
1017 return;
1018
David Rientjes255d0882014-02-10 14:25:39 -08001019 lockdep_assert_held(&n->list_lock);
Christoph Lameter643b1132007-05-06 14:49:42 -07001020 list_del(&page->lru);
Christoph Lameter643b1132007-05-06 14:49:42 -07001021}
1022
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001023/* Tracking of the number of slabs for debugging purposes */
1024static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1025{
1026 struct kmem_cache_node *n = get_node(s, node);
1027
1028 return atomic_long_read(&n->nr_slabs);
1029}
1030
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001031static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1032{
1033 return atomic_long_read(&n->nr_slabs);
1034}
1035
Christoph Lameter205ab992008-04-14 19:11:40 +03001036static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001037{
1038 struct kmem_cache_node *n = get_node(s, node);
1039
1040 /*
1041 * May be called early in order to allocate a slab for the
1042 * kmem_cache_node structure. Solve the chicken-egg
1043 * dilemma by deferring the increment of the count during
1044 * bootstrap (see early_kmem_cache_node_alloc).
1045 */
Joonsoo Kim338b2642013-01-21 17:01:27 +09001046 if (likely(n)) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001047 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001048 atomic_long_add(objects, &n->total_objects);
1049 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001050}
Christoph Lameter205ab992008-04-14 19:11:40 +03001051static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001052{
1053 struct kmem_cache_node *n = get_node(s, node);
1054
1055 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001056 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001057}
1058
1059/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001060static void setup_object_debug(struct kmem_cache *s, struct page *page,
1061 void *object)
1062{
1063 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1064 return;
1065
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001066 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter3ec09742007-05-16 22:11:00 -07001067 init_tracking(s, object);
1068}
1069
Chen Gangd0e0ac92013-07-15 09:05:29 +08001070static noinline int alloc_debug_processing(struct kmem_cache *s,
1071 struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001072 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001073{
1074 if (!check_slab(s, page))
1075 goto bad;
1076
Christoph Lameter81819f02007-05-06 14:49:36 -07001077 if (!check_valid_pointer(s, page, object)) {
1078 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -07001079 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001080 }
1081
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001082 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
Christoph Lameter81819f02007-05-06 14:49:36 -07001083 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001084
Christoph Lameter3ec09742007-05-16 22:11:00 -07001085 /* Success perform special debug activities for allocs */
1086 if (s->flags & SLAB_STORE_USER)
1087 set_track(s, object, TRACK_ALLOC, addr);
1088 trace(s, page, object, 1);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001089 init_object(s, object, SLUB_RED_ACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001090 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001091
Christoph Lameter81819f02007-05-06 14:49:36 -07001092bad:
1093 if (PageSlab(page)) {
1094 /*
1095 * If this is a slab page then lets do the best we can
1096 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -07001097 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001098 */
Christoph Lameter24922682007-07-17 04:03:18 -07001099 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +03001100 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001101 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001102 }
1103 return 0;
1104}
1105
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001106static noinline struct kmem_cache_node *free_debug_processing(
1107 struct kmem_cache *s, struct page *page, void *object,
1108 unsigned long addr, unsigned long *flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001109{
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001110 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001111
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001112 spin_lock_irqsave(&n->list_lock, *flags);
Christoph Lameter881db7f2011-06-01 12:25:53 -05001113 slab_lock(page);
1114
Christoph Lameter81819f02007-05-06 14:49:36 -07001115 if (!check_slab(s, page))
1116 goto fail;
1117
1118 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001119 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001120 goto fail;
1121 }
1122
1123 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -07001124 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -07001125 goto fail;
1126 }
1127
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001128 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001129 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07001130
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001131 if (unlikely(s != page->slab_cache)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001132 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001133 slab_err(s, page, "Attempt to free object(0x%p) "
1134 "outside of slab", object);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001135 } else if (!page->slab_cache) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001136 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -07001137 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -07001138 object);
Christoph Lameter70d71222007-05-06 14:49:47 -07001139 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -08001140 } else
Christoph Lameter24922682007-07-17 04:03:18 -07001141 object_err(s, page, object,
1142 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -07001143 goto fail;
1144 }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001145
Christoph Lameter3ec09742007-05-16 22:11:00 -07001146 if (s->flags & SLAB_STORE_USER)
1147 set_track(s, object, TRACK_FREE, addr);
1148 trace(s, page, object, 0);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001149 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001150out:
Christoph Lameter881db7f2011-06-01 12:25:53 -05001151 slab_unlock(page);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001152 /*
1153 * Keep node_lock to preserve integrity
1154 * until the object is actually freed
1155 */
1156 return n;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001157
Christoph Lameter81819f02007-05-06 14:49:36 -07001158fail:
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001159 slab_unlock(page);
1160 spin_unlock_irqrestore(&n->list_lock, *flags);
Christoph Lameter24922682007-07-17 04:03:18 -07001161 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001162 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001163}
1164
Christoph Lameter41ecc552007-05-09 02:32:44 -07001165static int __init setup_slub_debug(char *str)
1166{
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001167 slub_debug = DEBUG_DEFAULT_FLAGS;
1168 if (*str++ != '=' || !*str)
1169 /*
1170 * No options specified. Switch on full debugging.
1171 */
1172 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001173
1174 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001175 /*
1176 * No options but restriction on slabs. This means full
1177 * debugging for slabs matching a pattern.
1178 */
1179 goto check_slabs;
1180
David Rientjesfa5ec8a2009-07-07 00:14:14 -07001181 if (tolower(*str) == 'o') {
1182 /*
1183 * Avoid enabling debugging on caches if its minimum order
1184 * would increase as a result.
1185 */
1186 disable_higher_order_debug = 1;
1187 goto out;
1188 }
1189
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001190 slub_debug = 0;
1191 if (*str == '-')
1192 /*
1193 * Switch off all debugging measures.
1194 */
1195 goto out;
1196
1197 /*
1198 * Determine which debug features should be switched on
1199 */
Pekka Enberg06428782008-01-07 23:20:27 -08001200 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001201 switch (tolower(*str)) {
1202 case 'f':
1203 slub_debug |= SLAB_DEBUG_FREE;
1204 break;
1205 case 'z':
1206 slub_debug |= SLAB_RED_ZONE;
1207 break;
1208 case 'p':
1209 slub_debug |= SLAB_POISON;
1210 break;
1211 case 'u':
1212 slub_debug |= SLAB_STORE_USER;
1213 break;
1214 case 't':
1215 slub_debug |= SLAB_TRACE;
1216 break;
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03001217 case 'a':
1218 slub_debug |= SLAB_FAILSLAB;
1219 break;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001220 default:
1221 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001222 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001223 }
1224 }
1225
1226check_slabs:
1227 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001228 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001229out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001230 return 1;
1231}
1232
1233__setup("slub_debug", setup_slub_debug);
1234
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001235static unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001236 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001237 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001238{
1239 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001240 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001241 */
Christoph Lameterc6f58d92013-11-07 16:29:15 +00001242 if (slub_debug && (!slub_debug_slabs || (name &&
1243 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
David Rientjes3de47212009-07-27 18:30:35 -07001244 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001245
1246 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001247}
1248#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001249static inline void setup_object_debug(struct kmem_cache *s,
1250 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001251
Christoph Lameter3ec09742007-05-16 22:11:00 -07001252static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001253 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001254
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001255static inline struct kmem_cache_node *free_debug_processing(
1256 struct kmem_cache *s, struct page *page, void *object,
1257 unsigned long addr, unsigned long *flags) { return NULL; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001258
Christoph Lameter41ecc552007-05-09 02:32:44 -07001259static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1260 { return 1; }
1261static inline int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001262 void *object, u8 val) { return 1; }
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001263static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1264 struct page *page) {}
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001265static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1266 struct page *page) {}
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001267static inline unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001268 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001269 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001270{
1271 return flags;
1272}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001273#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001274
Ingo Molnarfdaa45e2009-09-15 11:00:26 +02001275#define disable_higher_order_debug 0
1276
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001277static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1278 { return 0; }
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001279static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1280 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001281static inline void inc_slabs_node(struct kmem_cache *s, int node,
1282 int objects) {}
1283static inline void dec_slabs_node(struct kmem_cache *s, int node,
1284 int objects) {}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001285
Roman Bobnievd56791b2013-10-08 15:58:57 -07001286static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1287{
1288 kmemleak_alloc(ptr, size, 1, flags);
1289}
1290
1291static inline void kfree_hook(const void *x)
1292{
1293 kmemleak_free(x);
1294}
1295
Christoph Lameter7d550c52010-08-25 14:07:16 -05001296static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1297 { return 0; }
1298
1299static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
Roman Bobnievd56791b2013-10-08 15:58:57 -07001300 void *object)
1301{
1302 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1303 flags & gfp_allowed_mask);
1304}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001305
Roman Bobnievd56791b2013-10-08 15:58:57 -07001306static inline void slab_free_hook(struct kmem_cache *s, void *x)
1307{
1308 kmemleak_free_recursive(x, s->flags);
1309}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001310
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05001311#endif /* CONFIG_SLUB_DEBUG */
Christoph Lameter205ab992008-04-14 19:11:40 +03001312
Christoph Lameter81819f02007-05-06 14:49:36 -07001313/*
1314 * Slab allocation and freeing
1315 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001316static inline struct page *alloc_slab_page(gfp_t flags, int node,
1317 struct kmem_cache_order_objects oo)
1318{
1319 int order = oo_order(oo);
1320
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001321 flags |= __GFP_NOTRACK;
1322
Christoph Lameter2154a332010-07-09 14:07:10 -05001323 if (node == NUMA_NO_NODE)
Christoph Lameter65c33762008-04-14 19:11:40 +03001324 return alloc_pages(flags, order);
1325 else
Minchan Kim6b65aaf2010-04-14 23:58:36 +09001326 return alloc_pages_exact_node(node, flags, order);
Christoph Lameter65c33762008-04-14 19:11:40 +03001327}
1328
Christoph Lameter81819f02007-05-06 14:49:36 -07001329static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1330{
Pekka Enberg06428782008-01-07 23:20:27 -08001331 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001332 struct kmem_cache_order_objects oo = s->oo;
Pekka Enbergba522702009-06-24 21:59:51 +03001333 gfp_t alloc_gfp;
Christoph Lameter81819f02007-05-06 14:49:36 -07001334
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001335 flags &= gfp_allowed_mask;
1336
1337 if (flags & __GFP_WAIT)
1338 local_irq_enable();
1339
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001340 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001341
Pekka Enbergba522702009-06-24 21:59:51 +03001342 /*
1343 * Let the initial higher-order allocation fail under memory pressure
1344 * so we fall-back to the minimum order allocation.
1345 */
1346 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1347
1348 page = alloc_slab_page(alloc_gfp, node, oo);
Christoph Lameter65c33762008-04-14 19:11:40 +03001349 if (unlikely(!page)) {
1350 oo = s->min;
1351 /*
1352 * Allocation may have failed due to fragmentation.
1353 * Try a lower order alloc if possible
1354 */
1355 page = alloc_slab_page(flags, node, oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001356
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001357 if (page)
1358 stat(s, ORDER_FALLBACK);
Christoph Lameter65c33762008-04-14 19:11:40 +03001359 }
Vegard Nossum5a896d92008-04-04 00:54:48 +02001360
David Rientjes737b7192012-07-09 14:00:38 -07001361 if (kmemcheck_enabled && page
Amerigo Wang5086c389c2009-08-19 21:44:13 +03001362 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001363 int pages = 1 << oo_order(oo);
1364
1365 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1366
1367 /*
1368 * Objects from caches that have a constructor don't get
1369 * cleared when they're allocated, so we need to do it here.
1370 */
1371 if (s->ctor)
1372 kmemcheck_mark_uninitialized_pages(page, pages);
1373 else
1374 kmemcheck_mark_unallocated_pages(page, pages);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001375 }
1376
David Rientjes737b7192012-07-09 14:00:38 -07001377 if (flags & __GFP_WAIT)
1378 local_irq_disable();
1379 if (!page)
1380 return NULL;
1381
Christoph Lameter834f3d12008-04-14 19:11:31 +03001382 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001383 mod_zone_page_state(page_zone(page),
1384 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1385 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001386 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001387
1388 return page;
1389}
1390
1391static void setup_object(struct kmem_cache *s, struct page *page,
1392 void *object)
1393{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001394 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001395 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001396 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001397}
1398
1399static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1400{
1401 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001402 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001403 void *last;
1404 void *p;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001405 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001406
Christoph Lameter6cb06222007-10-16 01:25:41 -07001407 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001408
Christoph Lameter6cb06222007-10-16 01:25:41 -07001409 page = allocate_slab(s,
1410 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001411 if (!page)
1412 goto out;
1413
Glauber Costa1f458cb2012-12-18 14:22:50 -08001414 order = compound_order(page);
Christoph Lameter205ab992008-04-14 19:11:40 +03001415 inc_slabs_node(s, page_to_nid(page), page->objects);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001416 memcg_bind_pages(s, order);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001417 page->slab_cache = s;
Joonsoo Kimc03f94c2012-05-18 00:47:47 +09001418 __SetPageSlab(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001419 if (page->pfmemalloc)
1420 SetPageSlabPfmemalloc(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001421
1422 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001423
1424 if (unlikely(s->flags & SLAB_POISON))
Glauber Costa1f458cb2012-12-18 14:22:50 -08001425 memset(start, POISON_INUSE, PAGE_SIZE << order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001426
1427 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001428 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001429 setup_object(s, page, last);
1430 set_freepointer(s, last, p);
1431 last = p;
1432 }
1433 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001434 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001435
1436 page->freelist = start;
Christoph Lametere6e82ea2011-08-09 16:12:24 -05001437 page->inuse = page->objects;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05001438 page->frozen = 1;
Christoph Lameter81819f02007-05-06 14:49:36 -07001439out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001440 return page;
1441}
1442
1443static void __free_slab(struct kmem_cache *s, struct page *page)
1444{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001445 int order = compound_order(page);
1446 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001447
Christoph Lameteraf537b02010-07-09 14:07:14 -05001448 if (kmem_cache_debug(s)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001449 void *p;
1450
1451 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001452 for_each_object(p, s, page_address(page),
1453 page->objects)
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001454 check_object(s, page, p, SLUB_RED_INACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001455 }
1456
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001457 kmemcheck_free_shadow(page, compound_order(page));
Vegard Nossum5a896d92008-04-04 00:54:48 +02001458
Christoph Lameter81819f02007-05-06 14:49:36 -07001459 mod_zone_page_state(page_zone(page),
1460 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1461 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001462 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001463
Mel Gorman072bb0a2012-07-31 16:43:58 -07001464 __ClearPageSlabPfmemalloc(page);
Christoph Lameter49bd5222008-04-14 18:52:18 +03001465 __ClearPageSlab(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001466
1467 memcg_release_pages(s, order);
Mel Gorman22b751c2013-02-22 16:34:59 -08001468 page_mapcount_reset(page);
Nick Piggin1eb5ac62009-05-05 19:13:44 +10001469 if (current->reclaim_state)
1470 current->reclaim_state->reclaimed_slab += pages;
Glauber Costad79923f2012-12-18 14:22:48 -08001471 __free_memcg_kmem_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001472}
1473
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001474#define need_reserve_slab_rcu \
1475 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1476
Christoph Lameter81819f02007-05-06 14:49:36 -07001477static void rcu_free_slab(struct rcu_head *h)
1478{
1479 struct page *page;
1480
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001481 if (need_reserve_slab_rcu)
1482 page = virt_to_head_page(h);
1483 else
1484 page = container_of((struct list_head *)h, struct page, lru);
1485
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001486 __free_slab(page->slab_cache, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001487}
1488
1489static void free_slab(struct kmem_cache *s, struct page *page)
1490{
1491 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001492 struct rcu_head *head;
1493
1494 if (need_reserve_slab_rcu) {
1495 int order = compound_order(page);
1496 int offset = (PAGE_SIZE << order) - s->reserved;
1497
1498 VM_BUG_ON(s->reserved != sizeof(*head));
1499 head = page_address(page) + offset;
1500 } else {
1501 /*
1502 * RCU free overloads the RCU head over the LRU
1503 */
1504 head = (void *)&page->lru;
1505 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001506
1507 call_rcu(head, rcu_free_slab);
1508 } else
1509 __free_slab(s, page);
1510}
1511
1512static void discard_slab(struct kmem_cache *s, struct page *page)
1513{
Christoph Lameter205ab992008-04-14 19:11:40 +03001514 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001515 free_slab(s, page);
1516}
1517
1518/*
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001519 * Management of partially allocated slabs.
Christoph Lameter81819f02007-05-06 14:49:36 -07001520 */
Steven Rostedt1e4dd942014-02-10 14:25:46 -08001521static inline void
1522__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001523{
Christoph Lametere95eed52007-05-06 14:49:44 -07001524 n->nr_partial++;
Shaohua Li136333d2011-08-24 08:57:52 +08001525 if (tail == DEACTIVATE_TO_TAIL)
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001526 list_add_tail(&page->lru, &n->partial);
1527 else
1528 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001529}
1530
Steven Rostedt1e4dd942014-02-10 14:25:46 -08001531static inline void add_partial(struct kmem_cache_node *n,
1532 struct page *page, int tail)
1533{
1534 lockdep_assert_held(&n->list_lock);
1535 __add_partial(n, page, tail);
1536}
1537
1538static inline void
1539__remove_partial(struct kmem_cache_node *n, struct page *page)
1540{
1541 list_del(&page->lru);
1542 n->nr_partial--;
1543}
1544
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001545static inline void remove_partial(struct kmem_cache_node *n,
Christoph Lameter62e346a2010-09-28 08:10:28 -05001546 struct page *page)
1547{
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001548 lockdep_assert_held(&n->list_lock);
Steven Rostedt1e4dd942014-02-10 14:25:46 -08001549 __remove_partial(n, page);
Christoph Lameter62e346a2010-09-28 08:10:28 -05001550}
1551
Christoph Lameter81819f02007-05-06 14:49:36 -07001552/*
Christoph Lameter7ced3712012-05-09 10:09:53 -05001553 * Remove slab from the partial list, freeze it and
1554 * return the pointer to the freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001555 *
Christoph Lameter497b66f2011-08-09 16:12:26 -05001556 * Returns a list of objects or NULL if it fails.
Christoph Lameter81819f02007-05-06 14:49:36 -07001557 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001558static inline void *acquire_slab(struct kmem_cache *s,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001559 struct kmem_cache_node *n, struct page *page,
Joonsoo Kim633b0762013-01-21 17:01:25 +09001560 int mode, int *objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001561{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001562 void *freelist;
1563 unsigned long counters;
1564 struct page new;
1565
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001566 lockdep_assert_held(&n->list_lock);
1567
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001568 /*
1569 * Zap the freelist and set the frozen bit.
1570 * The old freelist is the list of objects for the
1571 * per cpu allocation list.
1572 */
Christoph Lameter7ced3712012-05-09 10:09:53 -05001573 freelist = page->freelist;
1574 counters = page->counters;
1575 new.counters = counters;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001576 *objects = new.objects - new.inuse;
Pekka Enberg23910c52012-06-04 10:14:58 +03001577 if (mode) {
Christoph Lameter7ced3712012-05-09 10:09:53 -05001578 new.inuse = page->objects;
Pekka Enberg23910c52012-06-04 10:14:58 +03001579 new.freelist = NULL;
1580 } else {
1581 new.freelist = freelist;
1582 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001583
Dave Hansena0132ac2014-01-29 14:05:50 -08001584 VM_BUG_ON(new.frozen);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001585 new.frozen = 1;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001586
Christoph Lameter7ced3712012-05-09 10:09:53 -05001587 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001588 freelist, counters,
Joonsoo Kim02d76332012-05-17 00:13:02 +09001589 new.freelist, new.counters,
Christoph Lameter7ced3712012-05-09 10:09:53 -05001590 "acquire_slab"))
Christoph Lameter7ced3712012-05-09 10:09:53 -05001591 return NULL;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001592
1593 remove_partial(n, page);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001594 WARN_ON(!freelist);
Christoph Lameter49e22582011-08-09 16:12:27 -05001595 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001596}
1597
Joonsoo Kim633b0762013-01-21 17:01:25 +09001598static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001599static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
Christoph Lameter49e22582011-08-09 16:12:27 -05001600
Christoph Lameter81819f02007-05-06 14:49:36 -07001601/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001602 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001603 */
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001604static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1605 struct kmem_cache_cpu *c, gfp_t flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001606{
Christoph Lameter49e22582011-08-09 16:12:27 -05001607 struct page *page, *page2;
1608 void *object = NULL;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001609 int available = 0;
1610 int objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001611
1612 /*
1613 * Racy check. If we mistakenly see no partial slabs then we
1614 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001615 * partial slab and there is none available then get_partials()
1616 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001617 */
1618 if (!n || !n->nr_partial)
1619 return NULL;
1620
1621 spin_lock(&n->list_lock);
Christoph Lameter49e22582011-08-09 16:12:27 -05001622 list_for_each_entry_safe(page, page2, &n->partial, lru) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001623 void *t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001624
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001625 if (!pfmemalloc_match(page, flags))
1626 continue;
1627
Joonsoo Kim633b0762013-01-21 17:01:25 +09001628 t = acquire_slab(s, n, page, object == NULL, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05001629 if (!t)
1630 break;
1631
Joonsoo Kim633b0762013-01-21 17:01:25 +09001632 available += objects;
Alex,Shi12d79632011-09-07 10:26:36 +08001633 if (!object) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001634 c->page = page;
Christoph Lameter49e22582011-08-09 16:12:27 -05001635 stat(s, ALLOC_FROM_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001636 object = t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001637 } else {
Joonsoo Kim633b0762013-01-21 17:01:25 +09001638 put_cpu_partial(s, page, 0);
Alex Shi8028dce2012-02-03 23:34:56 +08001639 stat(s, CPU_PARTIAL_NODE);
Christoph Lameter49e22582011-08-09 16:12:27 -05001640 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001641 if (!kmem_cache_has_cpu_partial(s)
1642 || available > s->cpu_partial / 2)
Christoph Lameter49e22582011-08-09 16:12:27 -05001643 break;
1644
Christoph Lameter497b66f2011-08-09 16:12:26 -05001645 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001646 spin_unlock(&n->list_lock);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001647 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001648}
1649
1650/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001651 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001652 */
Joonsoo Kimde3ec032012-01-27 00:12:23 -08001653static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001654 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001655{
1656#ifdef CONFIG_NUMA
1657 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001658 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001659 struct zone *zone;
1660 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001661 void *object;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001662 unsigned int cpuset_mems_cookie;
Christoph Lameter81819f02007-05-06 14:49:36 -07001663
1664 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001665 * The defrag ratio allows a configuration of the tradeoffs between
1666 * inter node defragmentation and node local allocations. A lower
1667 * defrag_ratio increases the tendency to do local allocations
1668 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001669 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001670 * If the defrag_ratio is set to 0 then kmalloc() always
1671 * returns node local objects. If the ratio is higher then kmalloc()
1672 * may return off node objects because partial slabs are obtained
1673 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001674 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001675 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001676 * defrag_ratio = 1000) then every (well almost) allocation will
1677 * first attempt to defrag slab caches on other nodes. This means
1678 * scanning over all nodes to look for partial slabs which may be
1679 * expensive if we do it every time we are trying to find a slab
1680 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001681 */
Christoph Lameter98246012008-01-07 23:20:26 -08001682 if (!s->remote_node_defrag_ratio ||
1683 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001684 return NULL;
1685
Mel Gormancc9a6c82012-03-21 16:34:11 -07001686 do {
Mel Gormand26914d2014-04-03 14:47:24 -07001687 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07001688 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001689 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1690 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07001691
Mel Gormancc9a6c82012-03-21 16:34:11 -07001692 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001693
Mel Gormancc9a6c82012-03-21 16:34:11 -07001694 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1695 n->nr_partial > s->min_partial) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001696 object = get_partial_node(s, n, c, flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001697 if (object) {
1698 /*
Mel Gormand26914d2014-04-03 14:47:24 -07001699 * Don't check read_mems_allowed_retry()
1700 * here - if mems_allowed was updated in
1701 * parallel, that was a harmless race
1702 * between allocation and the cpuset
1703 * update
Mel Gormancc9a6c82012-03-21 16:34:11 -07001704 */
Mel Gormancc9a6c82012-03-21 16:34:11 -07001705 return object;
1706 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001707 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001708 }
Mel Gormand26914d2014-04-03 14:47:24 -07001709 } while (read_mems_allowed_retry(cpuset_mems_cookie));
Christoph Lameter81819f02007-05-06 14:49:36 -07001710#endif
1711 return NULL;
1712}
1713
1714/*
1715 * Get a partial page, lock it and return it.
1716 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001717static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001718 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001719{
Christoph Lameter497b66f2011-08-09 16:12:26 -05001720 void *object;
Christoph Lameter2154a332010-07-09 14:07:10 -05001721 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
Christoph Lameter81819f02007-05-06 14:49:36 -07001722
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001723 object = get_partial_node(s, get_node(s, searchnode), c, flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001724 if (object || node != NUMA_NO_NODE)
1725 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001726
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001727 return get_any_partial(s, flags, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001728}
1729
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001730#ifdef CONFIG_PREEMPT
1731/*
1732 * Calculate the next globally unique transaction for disambiguiation
1733 * during cmpxchg. The transactions start with the cpu number and are then
1734 * incremented by CONFIG_NR_CPUS.
1735 */
1736#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1737#else
1738/*
1739 * No preemption supported therefore also no need to check for
1740 * different cpus.
1741 */
1742#define TID_STEP 1
1743#endif
1744
1745static inline unsigned long next_tid(unsigned long tid)
1746{
1747 return tid + TID_STEP;
1748}
1749
1750static inline unsigned int tid_to_cpu(unsigned long tid)
1751{
1752 return tid % TID_STEP;
1753}
1754
1755static inline unsigned long tid_to_event(unsigned long tid)
1756{
1757 return tid / TID_STEP;
1758}
1759
1760static inline unsigned int init_tid(int cpu)
1761{
1762 return cpu;
1763}
1764
1765static inline void note_cmpxchg_failure(const char *n,
1766 const struct kmem_cache *s, unsigned long tid)
1767{
1768#ifdef SLUB_DEBUG_CMPXCHG
1769 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1770
1771 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1772
1773#ifdef CONFIG_PREEMPT
1774 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1775 printk("due to cpu change %d -> %d\n",
1776 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1777 else
1778#endif
1779 if (tid_to_event(tid) != tid_to_event(actual_tid))
1780 printk("due to cpu running other code. Event %ld->%ld\n",
1781 tid_to_event(tid), tid_to_event(actual_tid));
1782 else
1783 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1784 actual_tid, tid, next_tid(tid));
1785#endif
Christoph Lameter4fdccdf2011-03-22 13:35:00 -05001786 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001787}
1788
Fengguang Wu788e1aa2012-09-28 16:34:05 +08001789static void init_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001790{
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001791 int cpu;
1792
1793 for_each_possible_cpu(cpu)
1794 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001795}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001796
1797/*
1798 * Remove the cpu slab
1799 */
Chen Gangd0e0ac92013-07-15 09:05:29 +08001800static void deactivate_slab(struct kmem_cache *s, struct page *page,
1801 void *freelist)
Christoph Lameter81819f02007-05-06 14:49:36 -07001802{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001803 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001804 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1805 int lock = 0;
1806 enum slab_modes l = M_NONE, m = M_NONE;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001807 void *nextfree;
Shaohua Li136333d2011-08-24 08:57:52 +08001808 int tail = DEACTIVATE_TO_HEAD;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001809 struct page new;
1810 struct page old;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001811
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001812 if (page->freelist) {
Christoph Lameter84e554e62009-12-18 16:26:23 -06001813 stat(s, DEACTIVATE_REMOTE_FREES);
Shaohua Li136333d2011-08-24 08:57:52 +08001814 tail = DEACTIVATE_TO_TAIL;
Christoph Lameter894b8782007-05-10 03:15:16 -07001815 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001816
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001817 /*
1818 * Stage one: Free all available per cpu objects back
1819 * to the page freelist while it is still frozen. Leave the
1820 * last one.
1821 *
1822 * There is no need to take the list->lock because the page
1823 * is still frozen.
1824 */
1825 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1826 void *prior;
1827 unsigned long counters;
1828
1829 do {
1830 prior = page->freelist;
1831 counters = page->counters;
1832 set_freepointer(s, freelist, prior);
1833 new.counters = counters;
1834 new.inuse--;
Dave Hansena0132ac2014-01-29 14:05:50 -08001835 VM_BUG_ON(!new.frozen);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001836
Christoph Lameter1d071712011-07-14 12:49:12 -05001837 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001838 prior, counters,
1839 freelist, new.counters,
1840 "drain percpu freelist"));
1841
1842 freelist = nextfree;
1843 }
1844
1845 /*
1846 * Stage two: Ensure that the page is unfrozen while the
1847 * list presence reflects the actual number of objects
1848 * during unfreeze.
1849 *
1850 * We setup the list membership and then perform a cmpxchg
1851 * with the count. If there is a mismatch then the page
1852 * is not unfrozen but the page is on the wrong list.
1853 *
1854 * Then we restart the process which may have to remove
1855 * the page from the list that we just put it on again
1856 * because the number of objects in the slab may have
1857 * changed.
1858 */
1859redo:
1860
1861 old.freelist = page->freelist;
1862 old.counters = page->counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08001863 VM_BUG_ON(!old.frozen);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001864
1865 /* Determine target state of the slab */
1866 new.counters = old.counters;
1867 if (freelist) {
1868 new.inuse--;
1869 set_freepointer(s, freelist, old.freelist);
1870 new.freelist = freelist;
1871 } else
1872 new.freelist = old.freelist;
1873
1874 new.frozen = 0;
1875
Christoph Lameter81107182011-08-09 13:01:32 -05001876 if (!new.inuse && n->nr_partial > s->min_partial)
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001877 m = M_FREE;
1878 else if (new.freelist) {
1879 m = M_PARTIAL;
1880 if (!lock) {
1881 lock = 1;
1882 /*
1883 * Taking the spinlock removes the possiblity
1884 * that acquire_slab() will see a slab page that
1885 * is frozen
1886 */
1887 spin_lock(&n->list_lock);
1888 }
1889 } else {
1890 m = M_FULL;
1891 if (kmem_cache_debug(s) && !lock) {
1892 lock = 1;
1893 /*
1894 * This also ensures that the scanning of full
1895 * slabs from diagnostic functions will not see
1896 * any frozen slabs.
1897 */
1898 spin_lock(&n->list_lock);
1899 }
1900 }
1901
1902 if (l != m) {
1903
1904 if (l == M_PARTIAL)
1905
1906 remove_partial(n, page);
1907
1908 else if (l == M_FULL)
1909
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001910 remove_full(s, n, page);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001911
1912 if (m == M_PARTIAL) {
1913
1914 add_partial(n, page, tail);
Shaohua Li136333d2011-08-24 08:57:52 +08001915 stat(s, tail);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001916
1917 } else if (m == M_FULL) {
1918
1919 stat(s, DEACTIVATE_FULL);
1920 add_full(s, n, page);
1921
1922 }
1923 }
1924
1925 l = m;
Christoph Lameter1d071712011-07-14 12:49:12 -05001926 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001927 old.freelist, old.counters,
1928 new.freelist, new.counters,
1929 "unfreezing slab"))
1930 goto redo;
1931
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001932 if (lock)
1933 spin_unlock(&n->list_lock);
1934
1935 if (m == M_FREE) {
1936 stat(s, DEACTIVATE_EMPTY);
1937 discard_slab(s, page);
1938 stat(s, FREE_SLAB);
1939 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001940}
1941
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001942/*
1943 * Unfreeze all the cpu partial slabs.
1944 *
Christoph Lameter59a09912012-11-28 16:23:00 +00001945 * This function must be called with interrupts disabled
1946 * for the cpu using c (or some other guarantee must be there
1947 * to guarantee no concurrent accesses).
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001948 */
Christoph Lameter59a09912012-11-28 16:23:00 +00001949static void unfreeze_partials(struct kmem_cache *s,
1950 struct kmem_cache_cpu *c)
Christoph Lameter49e22582011-08-09 16:12:27 -05001951{
Joonsoo Kim345c9052013-06-19 14:05:52 +09001952#ifdef CONFIG_SLUB_CPU_PARTIAL
Joonsoo Kim43d77862012-06-09 02:23:16 +09001953 struct kmem_cache_node *n = NULL, *n2 = NULL;
Shaohua Li9ada1932011-11-14 13:34:13 +08001954 struct page *page, *discard_page = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05001955
1956 while ((page = c->partial)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001957 struct page new;
1958 struct page old;
1959
1960 c->partial = page->next;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001961
1962 n2 = get_node(s, page_to_nid(page));
1963 if (n != n2) {
1964 if (n)
1965 spin_unlock(&n->list_lock);
1966
1967 n = n2;
1968 spin_lock(&n->list_lock);
1969 }
Christoph Lameter49e22582011-08-09 16:12:27 -05001970
1971 do {
1972
1973 old.freelist = page->freelist;
1974 old.counters = page->counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08001975 VM_BUG_ON(!old.frozen);
Christoph Lameter49e22582011-08-09 16:12:27 -05001976
1977 new.counters = old.counters;
1978 new.freelist = old.freelist;
1979
1980 new.frozen = 0;
1981
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001982 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter49e22582011-08-09 16:12:27 -05001983 old.freelist, old.counters,
1984 new.freelist, new.counters,
1985 "unfreezing slab"));
1986
Joonsoo Kim43d77862012-06-09 02:23:16 +09001987 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
Shaohua Li9ada1932011-11-14 13:34:13 +08001988 page->next = discard_page;
1989 discard_page = page;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001990 } else {
1991 add_partial(n, page, DEACTIVATE_TO_TAIL);
1992 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001993 }
1994 }
1995
1996 if (n)
1997 spin_unlock(&n->list_lock);
Shaohua Li9ada1932011-11-14 13:34:13 +08001998
1999 while (discard_page) {
2000 page = discard_page;
2001 discard_page = discard_page->next;
2002
2003 stat(s, DEACTIVATE_EMPTY);
2004 discard_slab(s, page);
2005 stat(s, FREE_SLAB);
2006 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09002007#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002008}
2009
2010/*
2011 * Put a page that was just frozen (in __slab_free) into a partial page
2012 * slot if available. This is done without interrupts disabled and without
2013 * preemption disabled. The cmpxchg is racy and may put the partial page
2014 * onto a random cpus partial slot.
2015 *
2016 * If we did not find a slot then simply move all the partials to the
2017 * per node partial list.
2018 */
Joonsoo Kim633b0762013-01-21 17:01:25 +09002019static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
Christoph Lameter49e22582011-08-09 16:12:27 -05002020{
Joonsoo Kim345c9052013-06-19 14:05:52 +09002021#ifdef CONFIG_SLUB_CPU_PARTIAL
Christoph Lameter49e22582011-08-09 16:12:27 -05002022 struct page *oldpage;
2023 int pages;
2024 int pobjects;
2025
2026 do {
2027 pages = 0;
2028 pobjects = 0;
2029 oldpage = this_cpu_read(s->cpu_slab->partial);
2030
2031 if (oldpage) {
2032 pobjects = oldpage->pobjects;
2033 pages = oldpage->pages;
2034 if (drain && pobjects > s->cpu_partial) {
2035 unsigned long flags;
2036 /*
2037 * partial array is full. Move the existing
2038 * set to the per node partial list.
2039 */
2040 local_irq_save(flags);
Christoph Lameter59a09912012-11-28 16:23:00 +00002041 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
Christoph Lameter49e22582011-08-09 16:12:27 -05002042 local_irq_restore(flags);
Joonsoo Kime24fc412012-06-23 03:22:38 +09002043 oldpage = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05002044 pobjects = 0;
2045 pages = 0;
Alex Shi8028dce2012-02-03 23:34:56 +08002046 stat(s, CPU_PARTIAL_DRAIN);
Christoph Lameter49e22582011-08-09 16:12:27 -05002047 }
2048 }
2049
2050 pages++;
2051 pobjects += page->objects - page->inuse;
2052
2053 page->pages = pages;
2054 page->pobjects = pobjects;
2055 page->next = oldpage;
2056
Chen Gangd0e0ac92013-07-15 09:05:29 +08002057 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2058 != oldpage);
Joonsoo Kim345c9052013-06-19 14:05:52 +09002059#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002060}
2061
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002062static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002063{
Christoph Lameter84e554e62009-12-18 16:26:23 -06002064 stat(s, CPUSLAB_FLUSH);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002065 deactivate_slab(s, c->page, c->freelist);
2066
2067 c->tid = next_tid(c->tid);
2068 c->page = NULL;
2069 c->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002070}
2071
2072/*
2073 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08002074 *
Christoph Lameter81819f02007-05-06 14:49:36 -07002075 * Called from IPI handler with interrupts disabled.
2076 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002077static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07002078{
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002079 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07002080
Christoph Lameter49e22582011-08-09 16:12:27 -05002081 if (likely(c)) {
2082 if (c->page)
2083 flush_slab(s, c);
2084
Christoph Lameter59a09912012-11-28 16:23:00 +00002085 unfreeze_partials(s, c);
Christoph Lameter49e22582011-08-09 16:12:27 -05002086 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002087}
2088
2089static void flush_cpu_slab(void *d)
2090{
2091 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07002092
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002093 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07002094}
2095
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002096static bool has_cpu_slab(int cpu, void *info)
2097{
2098 struct kmem_cache *s = info;
2099 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2100
majianpeng02e1a9c2012-05-17 17:03:26 -07002101 return c->page || c->partial;
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002102}
2103
Christoph Lameter81819f02007-05-06 14:49:36 -07002104static void flush_all(struct kmem_cache *s)
2105{
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002106 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
Christoph Lameter81819f02007-05-06 14:49:36 -07002107}
2108
2109/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002110 * Check if the objects in a per cpu structure fit numa
2111 * locality expectations.
2112 */
Christoph Lameter57d437d2012-05-09 10:09:59 -05002113static inline int node_match(struct page *page, int node)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002114{
2115#ifdef CONFIG_NUMA
Christoph Lameter4d7868e2013-01-23 21:45:47 +00002116 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002117 return 0;
2118#endif
2119 return 1;
2120}
2121
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002122static int count_free(struct page *page)
2123{
2124 return page->objects - page->inuse;
2125}
2126
2127static unsigned long count_partial(struct kmem_cache_node *n,
2128 int (*get_count)(struct page *))
2129{
2130 unsigned long flags;
2131 unsigned long x = 0;
2132 struct page *page;
2133
2134 spin_lock_irqsave(&n->list_lock, flags);
2135 list_for_each_entry(page, &n->partial, lru)
2136 x += get_count(page);
2137 spin_unlock_irqrestore(&n->list_lock, flags);
2138 return x;
2139}
2140
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002141static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2142{
2143#ifdef CONFIG_SLUB_DEBUG
2144 return atomic_long_read(&n->total_objects);
2145#else
2146 return 0;
2147#endif
2148}
2149
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002150static noinline void
2151slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2152{
2153 int node;
2154
2155 printk(KERN_WARNING
2156 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2157 nid, gfpflags);
2158 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002159 "default order: %d, min order: %d\n", s->name, s->object_size,
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002160 s->size, oo_order(s->oo), oo_order(s->min));
2161
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002162 if (oo_order(s->min) > get_order(s->object_size))
David Rientjesfa5ec8a2009-07-07 00:14:14 -07002163 printk(KERN_WARNING " %s debugging increased min order, use "
2164 "slub_debug=O to disable.\n", s->name);
2165
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002166 for_each_online_node(node) {
2167 struct kmem_cache_node *n = get_node(s, node);
2168 unsigned long nr_slabs;
2169 unsigned long nr_objs;
2170 unsigned long nr_free;
2171
2172 if (!n)
2173 continue;
2174
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002175 nr_free = count_partial(n, count_free);
2176 nr_slabs = node_nr_slabs(n);
2177 nr_objs = node_nr_objs(n);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002178
2179 printk(KERN_WARNING
2180 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2181 node, nr_slabs, nr_objs, nr_free);
2182 }
2183}
2184
Christoph Lameter497b66f2011-08-09 16:12:26 -05002185static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2186 int node, struct kmem_cache_cpu **pc)
2187{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002188 void *freelist;
Christoph Lameter188fd062012-05-09 10:09:55 -05002189 struct kmem_cache_cpu *c = *pc;
2190 struct page *page;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002191
Christoph Lameter188fd062012-05-09 10:09:55 -05002192 freelist = get_partial(s, flags, node, c);
2193
2194 if (freelist)
2195 return freelist;
2196
2197 page = new_slab(s, flags, node);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002198 if (page) {
2199 c = __this_cpu_ptr(s->cpu_slab);
2200 if (c->page)
2201 flush_slab(s, c);
2202
2203 /*
2204 * No other reference to the page yet so we can
2205 * muck around with it freely without cmpxchg
2206 */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002207 freelist = page->freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002208 page->freelist = NULL;
2209
2210 stat(s, ALLOC_SLAB);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002211 c->page = page;
2212 *pc = c;
2213 } else
Christoph Lameter6faa6832012-05-09 10:09:51 -05002214 freelist = NULL;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002215
Christoph Lameter6faa6832012-05-09 10:09:51 -05002216 return freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002217}
2218
Mel Gorman072bb0a2012-07-31 16:43:58 -07002219static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2220{
2221 if (unlikely(PageSlabPfmemalloc(page)))
2222 return gfp_pfmemalloc_allowed(gfpflags);
2223
2224 return true;
2225}
2226
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002227/*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002228 * Check the page->freelist of a page and either transfer the freelist to the
2229 * per cpu freelist or deactivate the page.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002230 *
2231 * The page is still frozen if the return value is not NULL.
2232 *
2233 * If this function returns NULL then the page has been unfrozen.
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002234 *
2235 * This function must be called with interrupt disabled.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002236 */
2237static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2238{
2239 struct page new;
2240 unsigned long counters;
2241 void *freelist;
2242
2243 do {
2244 freelist = page->freelist;
2245 counters = page->counters;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002246
Christoph Lameter213eeb92011-11-11 14:07:14 -06002247 new.counters = counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08002248 VM_BUG_ON(!new.frozen);
Christoph Lameter213eeb92011-11-11 14:07:14 -06002249
2250 new.inuse = page->objects;
2251 new.frozen = freelist != NULL;
2252
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002253 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter213eeb92011-11-11 14:07:14 -06002254 freelist, counters,
2255 NULL, new.counters,
2256 "get_freelist"));
2257
2258 return freelist;
2259}
2260
2261/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002262 * Slow path. The lockless freelist is empty or we need to perform
2263 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07002264 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002265 * Processing is still very fast if new objects have been freed to the
2266 * regular freelist. In that case we simply take over the regular freelist
2267 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07002268 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002269 * If that is not working then we fall back to the partial lists. We take the
2270 * first element of the freelist as the object to allocate now and move the
2271 * rest of the freelist to the lockless freelist.
2272 *
2273 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08002274 * we need to allocate a new slab. This is the slowest path since it involves
2275 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07002276 */
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002277static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2278 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002279{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002280 void *freelist;
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002281 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002282 unsigned long flags;
2283
2284 local_irq_save(flags);
2285#ifdef CONFIG_PREEMPT
2286 /*
2287 * We may have been preempted and rescheduled on a different
2288 * cpu before disabling interrupts. Need to reload cpu area
2289 * pointer.
2290 */
2291 c = this_cpu_ptr(s->cpu_slab);
2292#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002293
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002294 page = c->page;
2295 if (!page)
Christoph Lameter81819f02007-05-06 14:49:36 -07002296 goto new_slab;
Christoph Lameter49e22582011-08-09 16:12:27 -05002297redo:
Christoph Lameter6faa6832012-05-09 10:09:51 -05002298
Christoph Lameter57d437d2012-05-09 10:09:59 -05002299 if (unlikely(!node_match(page, node))) {
Christoph Lametere36a2652011-06-01 12:25:57 -05002300 stat(s, ALLOC_NODE_MISMATCH);
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002301 deactivate_slab(s, page, c->freelist);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002302 c->page = NULL;
2303 c->freelist = NULL;
Christoph Lameterfc59c052011-06-01 12:25:56 -05002304 goto new_slab;
2305 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08002306
Mel Gorman072bb0a2012-07-31 16:43:58 -07002307 /*
2308 * By rights, we should be searching for a slab page that was
2309 * PFMEMALLOC but right now, we are losing the pfmemalloc
2310 * information when the page leaves the per-cpu allocator
2311 */
2312 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2313 deactivate_slab(s, page, c->freelist);
2314 c->page = NULL;
2315 c->freelist = NULL;
2316 goto new_slab;
2317 }
2318
Eric Dumazet73736e02011-12-13 04:57:06 +01002319 /* must check again c->freelist in case of cpu migration or IRQ */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002320 freelist = c->freelist;
2321 if (freelist)
Eric Dumazet73736e02011-12-13 04:57:06 +01002322 goto load_freelist;
2323
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002324 stat(s, ALLOC_SLOWPATH);
2325
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002326 freelist = get_freelist(s, page);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002327
Christoph Lameter6faa6832012-05-09 10:09:51 -05002328 if (!freelist) {
Christoph Lameter03e404a2011-06-01 12:25:58 -05002329 c->page = NULL;
2330 stat(s, DEACTIVATE_BYPASS);
Christoph Lameterfc59c052011-06-01 12:25:56 -05002331 goto new_slab;
Christoph Lameter03e404a2011-06-01 12:25:58 -05002332 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002333
Christoph Lameter81819f02007-05-06 14:49:36 -07002334 stat(s, ALLOC_REFILL);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08002335
Christoph Lameter894b8782007-05-10 03:15:16 -07002336load_freelist:
Christoph Lameter507effe2012-05-09 10:09:52 -05002337 /*
2338 * freelist is pointing to the list of objects to be used.
2339 * page is pointing to the page from which the objects are obtained.
2340 * That page must be frozen for per cpu allocations to work.
2341 */
Dave Hansena0132ac2014-01-29 14:05:50 -08002342 VM_BUG_ON(!c->page->frozen);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002343 c->freelist = get_freepointer(s, freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002344 c->tid = next_tid(c->tid);
2345 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002346 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07002347
Christoph Lameter81819f02007-05-06 14:49:36 -07002348new_slab:
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002349
Christoph Lameter49e22582011-08-09 16:12:27 -05002350 if (c->partial) {
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002351 page = c->page = c->partial;
2352 c->partial = page->next;
Christoph Lameter49e22582011-08-09 16:12:27 -05002353 stat(s, CPU_PARTIAL_ALLOC);
2354 c->freelist = NULL;
2355 goto redo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002356 }
2357
Christoph Lameter188fd062012-05-09 10:09:55 -05002358 freelist = new_slab_objects(s, gfpflags, node, &c);
Christoph Lameterb811c202007-10-16 23:25:51 -07002359
Christoph Lameterf46974362012-05-09 10:09:54 -05002360 if (unlikely(!freelist)) {
2361 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2362 slab_out_of_memory(s, gfpflags, node);
Christoph Lameter01ad8a72011-04-15 14:48:14 -05002363
Christoph Lameterf46974362012-05-09 10:09:54 -05002364 local_irq_restore(flags);
2365 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002366 }
Christoph Lameter894b8782007-05-10 03:15:16 -07002367
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002368 page = c->page;
Christoph Lameter5091b742012-07-31 16:44:00 -07002369 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002370 goto load_freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002371
Christoph Lameter497b66f2011-08-09 16:12:26 -05002372 /* Only entered in the debug case */
Chen Gangd0e0ac92013-07-15 09:05:29 +08002373 if (kmem_cache_debug(s) &&
2374 !alloc_debug_processing(s, page, freelist, addr))
Christoph Lameter497b66f2011-08-09 16:12:26 -05002375 goto new_slab; /* Slab failed checks. Next slab needed */
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002376
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002377 deactivate_slab(s, page, get_freepointer(s, freelist));
Christoph Lameterc17dda42012-05-09 10:09:57 -05002378 c->page = NULL;
2379 c->freelist = NULL;
Christoph Lametera71ae472011-05-25 09:47:43 -05002380 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002381 return freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002382}
2383
2384/*
2385 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2386 * have the fastpath folded into their functions. So no function call
2387 * overhead for requests that can be satisfied on the fastpath.
2388 *
2389 * The fastpath works by first checking if the lockless freelist can be used.
2390 * If not then __slab_alloc is called for slow processing.
2391 *
2392 * Otherwise we can simply pick the next object from the lockless free list.
2393 */
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002394static __always_inline void *slab_alloc_node(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002395 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002396{
Christoph Lameter894b8782007-05-10 03:15:16 -07002397 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002398 struct kmem_cache_cpu *c;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002399 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002400 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002401
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002402 if (slab_pre_alloc_hook(s, gfpflags))
Akinobu Mita773ff602008-12-23 19:37:01 +09002403 return NULL;
2404
Glauber Costad79923f2012-12-18 14:22:48 -08002405 s = memcg_kmem_get_cache(s, gfpflags);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002406redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002407 /*
2408 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2409 * enabled. We may switch back and forth between cpus while
2410 * reading from one cpu area. That does not matter as long
2411 * as we end up on the original cpu again when doing the cmpxchg.
Christoph Lameter7cccd802013-01-23 21:45:48 +00002412 *
2413 * Preemption is disabled for the retrieval of the tid because that
2414 * must occur from the current processor. We cannot allow rescheduling
2415 * on a different processor between the determination of the pointer
2416 * and the retrieval of the tid.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002417 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002418 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002419 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002420
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002421 /*
2422 * The transaction ids are globally unique per cpu and per operation on
2423 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2424 * occurs on the right processor and that there was no operation on the
2425 * linked list in between.
2426 */
2427 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002428 preempt_enable();
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002429
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002430 object = c->freelist;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002431 page = c->page;
Libinac6434e2013-07-18 15:39:51 +08002432 if (unlikely(!object || !node_match(page, node)))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002433 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07002434
2435 else {
Eric Dumazet0ad95002011-12-16 16:25:34 +01002436 void *next_object = get_freepointer_safe(s, object);
2437
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002438 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002439 * The cmpxchg will only match if there was no additional
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002440 * operation and if we are on the right processor.
2441 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002442 * The cmpxchg does the following atomically (without lock
2443 * semantics!)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002444 * 1. Relocate first pointer to the current per cpu area.
2445 * 2. Verify that tid and freelist have not been changed
2446 * 3. If they were not changed replace tid and freelist
2447 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002448 * Since this is without lock semantics the protection is only
2449 * against code executing on this cpu *not* from access by
2450 * other cpus.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002451 */
Christoph Lameter933393f2011-12-22 11:58:51 -06002452 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002453 s->cpu_slab->freelist, s->cpu_slab->tid,
2454 object, tid,
Eric Dumazet0ad95002011-12-16 16:25:34 +01002455 next_object, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002456
2457 note_cmpxchg_failure("slab_alloc", s, tid);
2458 goto redo;
2459 }
Eric Dumazet0ad95002011-12-16 16:25:34 +01002460 prefetch_freepointer(s, next_object);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002461 stat(s, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002462 }
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002463
Pekka Enberg74e21342009-11-25 20:14:48 +02002464 if (unlikely(gfpflags & __GFP_ZERO) && object)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002465 memset(object, 0, s->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07002466
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002467 slab_post_alloc_hook(s, gfpflags, object);
Vegard Nossum5a896d92008-04-04 00:54:48 +02002468
Christoph Lameter894b8782007-05-10 03:15:16 -07002469 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07002470}
2471
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002472static __always_inline void *slab_alloc(struct kmem_cache *s,
2473 gfp_t gfpflags, unsigned long addr)
2474{
2475 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2476}
2477
Christoph Lameter81819f02007-05-06 14:49:36 -07002478void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2479{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002480 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002481
Chen Gangd0e0ac92013-07-15 09:05:29 +08002482 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2483 s->size, gfpflags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002484
2485 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002486}
2487EXPORT_SYMBOL(kmem_cache_alloc);
2488
Li Zefan0f24f122009-12-11 15:45:30 +08002489#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002490void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002491{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002492 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002493 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2494 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002495}
Richard Kennedy4a923792010-10-21 10:29:19 +01002496EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002497#endif
2498
Christoph Lameter81819f02007-05-06 14:49:36 -07002499#ifdef CONFIG_NUMA
2500void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2501{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002502 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002503
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002504 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002505 s->object_size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002506
2507 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002508}
2509EXPORT_SYMBOL(kmem_cache_alloc_node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002510
Li Zefan0f24f122009-12-11 15:45:30 +08002511#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002512void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002513 gfp_t gfpflags,
Richard Kennedy4a923792010-10-21 10:29:19 +01002514 int node, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002515{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002516 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002517
2518 trace_kmalloc_node(_RET_IP_, ret,
2519 size, s->size, gfpflags, node);
2520 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002521}
Richard Kennedy4a923792010-10-21 10:29:19 +01002522EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002523#endif
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09002524#endif
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002525
Christoph Lameter81819f02007-05-06 14:49:36 -07002526/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002527 * Slow patch handling. This may still be called frequently since objects
2528 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07002529 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002530 * So we still attempt to reduce cache line usage. Just take the slab
2531 * lock and free the item. If there is no additional partial page
2532 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07002533 */
Christoph Lameter894b8782007-05-10 03:15:16 -07002534static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterff120592009-12-18 16:26:22 -06002535 void *x, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07002536{
2537 void *prior;
2538 void **object = (void *)x;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002539 int was_frozen;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002540 struct page new;
2541 unsigned long counters;
2542 struct kmem_cache_node *n = NULL;
Christoph Lameter61728d12011-06-01 12:25:51 -05002543 unsigned long uninitialized_var(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002544
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002545 stat(s, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07002546
Christoph Lameter19c7ff92012-05-30 12:54:46 -05002547 if (kmem_cache_debug(s) &&
2548 !(n = free_debug_processing(s, page, x, addr, &flags)))
Christoph Lameter80f08c12011-06-01 12:25:55 -05002549 return;
Christoph Lameter6446faa2008-02-15 23:45:26 -08002550
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002551 do {
Joonsoo Kim837d6782012-08-16 00:02:40 +09002552 if (unlikely(n)) {
2553 spin_unlock_irqrestore(&n->list_lock, flags);
2554 n = NULL;
2555 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002556 prior = page->freelist;
2557 counters = page->counters;
2558 set_freepointer(s, object, prior);
2559 new.counters = counters;
2560 was_frozen = new.frozen;
2561 new.inuse--;
Joonsoo Kim837d6782012-08-16 00:02:40 +09002562 if ((!new.inuse || !prior) && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002563
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002564 if (kmem_cache_has_cpu_partial(s) && !prior) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002565
2566 /*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002567 * Slab was on no list before and will be
2568 * partially empty
2569 * We can defer the list move and instead
2570 * freeze it.
Christoph Lameter49e22582011-08-09 16:12:27 -05002571 */
2572 new.frozen = 1;
2573
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002574 } else { /* Needs to be taken off a list */
Christoph Lameter49e22582011-08-09 16:12:27 -05002575
2576 n = get_node(s, page_to_nid(page));
2577 /*
2578 * Speculatively acquire the list_lock.
2579 * If the cmpxchg does not succeed then we may
2580 * drop the list_lock without any processing.
2581 *
2582 * Otherwise the list_lock will synchronize with
2583 * other processors updating the list of slabs.
2584 */
2585 spin_lock_irqsave(&n->list_lock, flags);
2586
2587 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002588 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002589
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002590 } while (!cmpxchg_double_slab(s, page,
2591 prior, counters,
2592 object, new.counters,
2593 "__slab_free"));
Christoph Lameter81819f02007-05-06 14:49:36 -07002594
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002595 if (likely(!n)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002596
2597 /*
2598 * If we just froze the page then put it onto the
2599 * per cpu partial list.
2600 */
Alex Shi8028dce2012-02-03 23:34:56 +08002601 if (new.frozen && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002602 put_cpu_partial(s, page, 1);
Alex Shi8028dce2012-02-03 23:34:56 +08002603 stat(s, CPU_PARTIAL_FREE);
2604 }
Christoph Lameter49e22582011-08-09 16:12:27 -05002605 /*
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002606 * The list lock was not taken therefore no list
2607 * activity can be necessary.
2608 */
2609 if (was_frozen)
2610 stat(s, FREE_FROZEN);
Christoph Lameter80f08c12011-06-01 12:25:55 -05002611 return;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002612 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002613
Joonsoo Kim837d6782012-08-16 00:02:40 +09002614 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2615 goto slab_empty;
Christoph Lameter81819f02007-05-06 14:49:36 -07002616
Joonsoo Kim837d6782012-08-16 00:02:40 +09002617 /*
2618 * Objects left in the slab. If it was not on the partial list before
2619 * then add it.
2620 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09002621 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2622 if (kmem_cache_debug(s))
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002623 remove_full(s, n, page);
Joonsoo Kim837d6782012-08-16 00:02:40 +09002624 add_partial(n, page, DEACTIVATE_TO_TAIL);
2625 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002626 }
Christoph Lameter80f08c12011-06-01 12:25:55 -05002627 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002628 return;
2629
2630slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08002631 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002632 /*
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002633 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07002634 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05002635 remove_partial(n, page);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002636 stat(s, FREE_REMOVE_PARTIAL);
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002637 } else {
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002638 /* Slab must be on the full list */
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002639 remove_full(s, n, page);
2640 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002641
Christoph Lameter80f08c12011-06-01 12:25:55 -05002642 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002643 stat(s, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07002644 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002645}
2646
Christoph Lameter894b8782007-05-10 03:15:16 -07002647/*
2648 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2649 * can perform fastpath freeing without additional function calls.
2650 *
2651 * The fastpath is only possible if we are freeing to the current cpu slab
2652 * of this processor. This typically the case if we have just allocated
2653 * the item before.
2654 *
2655 * If fastpath is not possible then fall back to __slab_free where we deal
2656 * with all sorts of special processing.
2657 */
Pekka Enberg06428782008-01-07 23:20:27 -08002658static __always_inline void slab_free(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002659 struct page *page, void *x, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002660{
2661 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002662 struct kmem_cache_cpu *c;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002663 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002664
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002665 slab_free_hook(s, x);
2666
Christoph Lametera24c5a02011-03-15 12:45:21 -05002667redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002668 /*
2669 * Determine the currently cpus per cpu slab.
2670 * The cpu may change afterward. However that does not matter since
2671 * data is retrieved via this pointer. If we are on the same cpu
2672 * during the cmpxchg then the free will succedd.
2673 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002674 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002675 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002676
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002677 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002678 preempt_enable();
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002679
Christoph Lameter442b06b2011-05-17 16:29:31 -05002680 if (likely(page == c->page)) {
Christoph Lameterff120592009-12-18 16:26:22 -06002681 set_freepointer(s, object, c->freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002682
Christoph Lameter933393f2011-12-22 11:58:51 -06002683 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002684 s->cpu_slab->freelist, s->cpu_slab->tid,
2685 c->freelist, tid,
2686 object, next_tid(tid)))) {
2687
2688 note_cmpxchg_failure("slab_free", s, tid);
2689 goto redo;
2690 }
Christoph Lameter84e554e62009-12-18 16:26:23 -06002691 stat(s, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002692 } else
Christoph Lameterff120592009-12-18 16:26:22 -06002693 __slab_free(s, page, x, addr);
Christoph Lameter894b8782007-05-10 03:15:16 -07002694
Christoph Lameter894b8782007-05-10 03:15:16 -07002695}
2696
Christoph Lameter81819f02007-05-06 14:49:36 -07002697void kmem_cache_free(struct kmem_cache *s, void *x)
2698{
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002699 s = cache_from_obj(s, x);
2700 if (!s)
Christoph Lameter79576102012-09-04 23:06:14 +00002701 return;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002702 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002703 trace_kmem_cache_free(_RET_IP_, x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002704}
2705EXPORT_SYMBOL(kmem_cache_free);
2706
Christoph Lameter81819f02007-05-06 14:49:36 -07002707/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002708 * Object placement in a slab is made very easy because we always start at
2709 * offset 0. If we tune the size of the object to the alignment then we can
2710 * get the required alignment by putting one properly sized object after
2711 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07002712 *
2713 * Notice that the allocation order determines the sizes of the per cpu
2714 * caches. Each processor has always one slab available for allocations.
2715 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07002716 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07002717 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07002718 */
2719
2720/*
2721 * Mininum / Maximum order of slab pages. This influences locking overhead
2722 * and slab fragmentation. A higher order reduces the number of partial slabs
2723 * and increases the number of allocations possible without having to
2724 * take the list_lock.
2725 */
2726static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03002727static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002728static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002729
2730/*
2731 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07002732 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07002733 */
2734static int slub_nomerge;
2735
2736/*
Christoph Lameter81819f02007-05-06 14:49:36 -07002737 * Calculate the order of allocation given an slab object size.
2738 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002739 * The order of allocation has significant impact on performance and other
2740 * system components. Generally order 0 allocations should be preferred since
2741 * order 0 does not cause fragmentation in the page allocator. Larger objects
2742 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002743 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07002744 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07002745 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002746 * In order to reach satisfactory performance we must ensure that a minimum
2747 * number of objects is in one slab. Otherwise we may generate too much
2748 * activity on the partial lists which requires taking the list_lock. This is
2749 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07002750 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002751 * slub_max_order specifies the order where we begin to stop considering the
2752 * number of objects in a slab as critical. If we reach slub_max_order then
2753 * we try to keep the page order as low as possible. So we accept more waste
2754 * of space in favor of a small page order.
2755 *
2756 * Higher order allocations also allow the placement of more objects in a
2757 * slab and thereby reduce object handling overhead. If the user has
2758 * requested a higher mininum order then we start with that one instead of
2759 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07002760 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002761static inline int slab_order(int size, int min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002762 int max_order, int fract_leftover, int reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002763{
2764 int order;
2765 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07002766 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002767
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002768 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04002769 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03002770
Christoph Lameter6300ea72007-07-17 04:03:20 -07002771 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002772 fls(min_objects * size - 1) - PAGE_SHIFT);
2773 order <= max_order; order++) {
2774
Christoph Lameter81819f02007-05-06 14:49:36 -07002775 unsigned long slab_size = PAGE_SIZE << order;
2776
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002777 if (slab_size < min_objects * size + reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002778 continue;
2779
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002780 rem = (slab_size - reserved) % size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002781
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002782 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07002783 break;
2784
2785 }
Christoph Lameter672bba32007-05-09 02:32:39 -07002786
Christoph Lameter81819f02007-05-06 14:49:36 -07002787 return order;
2788}
2789
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002790static inline int calculate_order(int size, int reserved)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002791{
2792 int order;
2793 int min_objects;
2794 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002795 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002796
2797 /*
2798 * Attempt to find best configuration for a slab. This
2799 * works by first attempting to generate a layout with
2800 * the best configuration and backing off gradually.
2801 *
2802 * First we reduce the acceptable waste in a slab. Then
2803 * we reduce the minimum objects required in a slab.
2804 */
2805 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002806 if (!min_objects)
2807 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002808 max_objects = order_objects(slub_max_order, size, reserved);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002809 min_objects = min(min_objects, max_objects);
2810
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002811 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002812 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002813 while (fraction >= 4) {
2814 order = slab_order(size, min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002815 slub_max_order, fraction, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002816 if (order <= slub_max_order)
2817 return order;
2818 fraction /= 2;
2819 }
Amerigo Wang5086c389c2009-08-19 21:44:13 +03002820 min_objects--;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002821 }
2822
2823 /*
2824 * We were unable to place multiple objects in a slab. Now
2825 * lets see if we can place a single object there.
2826 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002827 order = slab_order(size, 1, slub_max_order, 1, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002828 if (order <= slub_max_order)
2829 return order;
2830
2831 /*
2832 * Doh this slab cannot be placed using slub_max_order.
2833 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002834 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
David Rientjes818cf592009-04-23 09:58:22 +03002835 if (order < MAX_ORDER)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002836 return order;
2837 return -ENOSYS;
2838}
2839
Pekka Enberg5595cff2008-08-05 09:28:47 +03002840static void
Joonsoo Kim40534972012-05-11 00:50:47 +09002841init_kmem_cache_node(struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002842{
2843 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002844 spin_lock_init(&n->list_lock);
2845 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002846#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002847 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07002848 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07002849 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002850#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002851}
2852
Christoph Lameter55136592010-08-20 12:37:13 -05002853static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002854{
Christoph Lameter6c182dc2010-08-20 12:37:14 -05002855 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
Christoph Lameter95a05b42013-01-10 19:14:19 +00002856 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002857
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002858 /*
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002859 * Must align to double word boundary for the double cmpxchg
2860 * instructions to work; see __pcpu_double_call_return_bool().
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002861 */
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002862 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2863 2 * sizeof(void *));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002864
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002865 if (!s->cpu_slab)
2866 return 0;
2867
2868 init_kmem_cache_cpus(s);
2869
2870 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002871}
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002872
Christoph Lameter51df1142010-08-20 12:37:15 -05002873static struct kmem_cache *kmem_cache_node;
2874
Christoph Lameter81819f02007-05-06 14:49:36 -07002875/*
2876 * No kmalloc_node yet so do it by hand. We know that this is the first
2877 * slab on the node for this slabcache. There are no concurrent accesses
2878 * possible.
2879 *
Zhi Yong Wu721ae222013-11-08 20:47:37 +08002880 * Note that this function only works on the kmem_cache_node
2881 * when allocating for the kmem_cache_node. This is used for bootstrapping
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002882 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002883 */
Christoph Lameter55136592010-08-20 12:37:13 -05002884static void early_kmem_cache_node_alloc(int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002885{
2886 struct page *page;
2887 struct kmem_cache_node *n;
2888
Christoph Lameter51df1142010-08-20 12:37:15 -05002889 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
Christoph Lameter81819f02007-05-06 14:49:36 -07002890
Christoph Lameter51df1142010-08-20 12:37:15 -05002891 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002892
2893 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002894 if (page_to_nid(page) != node) {
2895 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2896 "node %d\n", node);
2897 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2898 "in order to be able to continue\n");
2899 }
2900
Christoph Lameter81819f02007-05-06 14:49:36 -07002901 n = page->freelist;
2902 BUG_ON(!n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002903 page->freelist = get_freepointer(kmem_cache_node, n);
Christoph Lametere6e82ea2011-08-09 16:12:24 -05002904 page->inuse = 1;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05002905 page->frozen = 0;
Christoph Lameter51df1142010-08-20 12:37:15 -05002906 kmem_cache_node->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002907#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterf7cb1932010-09-29 07:15:01 -05002908 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
Christoph Lameter51df1142010-08-20 12:37:15 -05002909 init_tracking(kmem_cache_node, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002910#endif
Joonsoo Kim40534972012-05-11 00:50:47 +09002911 init_kmem_cache_node(n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002912 inc_slabs_node(kmem_cache_node, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002913
Dave Hansen67b6c902014-01-24 07:20:23 -08002914 /*
Steven Rostedt1e4dd942014-02-10 14:25:46 -08002915 * No locks need to be taken here as it has just been
2916 * initialized and there is no concurrent access.
Dave Hansen67b6c902014-01-24 07:20:23 -08002917 */
Steven Rostedt1e4dd942014-02-10 14:25:46 -08002918 __add_partial(n, page, DEACTIVATE_TO_HEAD);
Christoph Lameter81819f02007-05-06 14:49:36 -07002919}
2920
2921static void free_kmem_cache_nodes(struct kmem_cache *s)
2922{
2923 int node;
2924
Christoph Lameterf64dc582007-10-16 01:25:33 -07002925 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002926 struct kmem_cache_node *n = s->node[node];
Christoph Lameter51df1142010-08-20 12:37:15 -05002927
Alexander Duyck73367bd2010-05-21 14:41:35 -07002928 if (n)
Christoph Lameter51df1142010-08-20 12:37:15 -05002929 kmem_cache_free(kmem_cache_node, n);
2930
Christoph Lameter81819f02007-05-06 14:49:36 -07002931 s->node[node] = NULL;
2932 }
2933}
2934
Christoph Lameter55136592010-08-20 12:37:13 -05002935static int init_kmem_cache_nodes(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002936{
2937 int node;
Christoph Lameter81819f02007-05-06 14:49:36 -07002938
Christoph Lameterf64dc582007-10-16 01:25:33 -07002939 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002940 struct kmem_cache_node *n;
2941
Alexander Duyck73367bd2010-05-21 14:41:35 -07002942 if (slab_state == DOWN) {
Christoph Lameter55136592010-08-20 12:37:13 -05002943 early_kmem_cache_node_alloc(node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002944 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07002945 }
Christoph Lameter51df1142010-08-20 12:37:15 -05002946 n = kmem_cache_alloc_node(kmem_cache_node,
Christoph Lameter55136592010-08-20 12:37:13 -05002947 GFP_KERNEL, node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002948
2949 if (!n) {
2950 free_kmem_cache_nodes(s);
2951 return 0;
2952 }
2953
Christoph Lameter81819f02007-05-06 14:49:36 -07002954 s->node[node] = n;
Joonsoo Kim40534972012-05-11 00:50:47 +09002955 init_kmem_cache_node(n);
Christoph Lameter81819f02007-05-06 14:49:36 -07002956 }
2957 return 1;
2958}
Christoph Lameter81819f02007-05-06 14:49:36 -07002959
David Rientjesc0bdb232009-02-25 09:16:35 +02002960static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08002961{
2962 if (min < MIN_PARTIAL)
2963 min = MIN_PARTIAL;
2964 else if (min > MAX_PARTIAL)
2965 min = MAX_PARTIAL;
2966 s->min_partial = min;
2967}
2968
Christoph Lameter81819f02007-05-06 14:49:36 -07002969/*
2970 * calculate_sizes() determines the order and the distribution of data within
2971 * a slab object.
2972 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002973static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002974{
2975 unsigned long flags = s->flags;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002976 unsigned long size = s->object_size;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002977 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002978
2979 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002980 * Round up object size to the next word boundary. We can only
2981 * place the free pointer at word boundaries and this determines
2982 * the possible location of the free pointer.
2983 */
2984 size = ALIGN(size, sizeof(void *));
2985
2986#ifdef CONFIG_SLUB_DEBUG
2987 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002988 * Determine if we can poison the object itself. If the user of
2989 * the slab may touch the object after free or before allocation
2990 * then we should never poison the object itself.
2991 */
2992 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002993 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002994 s->flags |= __OBJECT_POISON;
2995 else
2996 s->flags &= ~__OBJECT_POISON;
2997
Christoph Lameter81819f02007-05-06 14:49:36 -07002998
2999 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003000 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07003001 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07003002 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07003003 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003004 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003005 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07003006#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003007
3008 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003009 * With that we have determined the number of bytes in actual use
3010 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07003011 */
3012 s->inuse = size;
3013
3014 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07003015 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003016 /*
3017 * Relocate free pointer after the object if it is not
3018 * permitted to overwrite the first word of the object on
3019 * kmem_cache_free.
3020 *
3021 * This is the case if we do RCU, have a constructor or
3022 * destructor or are poisoning the objects.
3023 */
3024 s->offset = size;
3025 size += sizeof(void *);
3026 }
3027
Christoph Lameterc12b3c62007-05-23 13:57:31 -07003028#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07003029 if (flags & SLAB_STORE_USER)
3030 /*
3031 * Need to store information about allocs and frees after
3032 * the object.
3033 */
3034 size += 2 * sizeof(struct track);
3035
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07003036 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07003037 /*
3038 * Add some empty padding so that we can catch
3039 * overwrites from earlier objects rather than let
3040 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01003041 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07003042 * of the object.
3043 */
3044 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07003045#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07003046
Christoph Lameter81819f02007-05-06 14:49:36 -07003047 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07003048 * SLUB stores one object immediately after another beginning from
3049 * offset 0. In order to align the objects we have to simply size
3050 * each object to conform to the alignment.
3051 */
Christoph Lameter45906852012-11-28 16:23:16 +00003052 size = ALIGN(size, s->align);
Christoph Lameter81819f02007-05-06 14:49:36 -07003053 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003054 if (forced_order >= 0)
3055 order = forced_order;
3056 else
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003057 order = calculate_order(size, s->reserved);
Christoph Lameter81819f02007-05-06 14:49:36 -07003058
Christoph Lameter834f3d12008-04-14 19:11:31 +03003059 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07003060 return 0;
3061
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003062 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003063 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003064 s->allocflags |= __GFP_COMP;
3065
3066 if (s->flags & SLAB_CACHE_DMA)
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003067 s->allocflags |= GFP_DMA;
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003068
3069 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3070 s->allocflags |= __GFP_RECLAIMABLE;
3071
Christoph Lameter81819f02007-05-06 14:49:36 -07003072 /*
3073 * Determine the number of objects per slab
3074 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003075 s->oo = oo_make(order, size, s->reserved);
3076 s->min = oo_make(get_order(size), size, s->reserved);
Christoph Lameter205ab992008-04-14 19:11:40 +03003077 if (oo_objects(s->oo) > oo_objects(s->max))
3078 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07003079
Christoph Lameter834f3d12008-04-14 19:11:31 +03003080 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07003081}
3082
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003083static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003084{
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003085 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003086 s->reserved = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003087
Lai Jiangshanda9a6382011-03-10 15:22:00 +08003088 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3089 s->reserved = sizeof(struct rcu_head);
Christoph Lameter81819f02007-05-06 14:49:36 -07003090
Christoph Lameter06b285d2008-04-14 19:11:41 +03003091 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07003092 goto error;
David Rientjes3de47212009-07-27 18:30:35 -07003093 if (disable_higher_order_debug) {
3094 /*
3095 * Disable debugging flags that store metadata if the min slab
3096 * order increased.
3097 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003098 if (get_order(s->size) > get_order(s->object_size)) {
David Rientjes3de47212009-07-27 18:30:35 -07003099 s->flags &= ~DEBUG_METADATA_FLAGS;
3100 s->offset = 0;
3101 if (!calculate_sizes(s, -1))
3102 goto error;
3103 }
3104 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003105
Heiko Carstens25654092012-01-12 17:17:33 -08003106#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3107 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -05003108 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3109 /* Enable fast mode */
3110 s->flags |= __CMPXCHG_DOUBLE;
3111#endif
3112
David Rientjes3b89d7d2009-02-22 17:40:07 -08003113 /*
3114 * The larger the object size is, the more pages we want on the partial
3115 * list to avoid pounding the page allocator excessively.
3116 */
Christoph Lameter49e22582011-08-09 16:12:27 -05003117 set_min_partial(s, ilog2(s->size) / 2);
3118
3119 /*
3120 * cpu_partial determined the maximum number of objects kept in the
3121 * per cpu partial lists of a processor.
3122 *
3123 * Per cpu partial lists mainly contain slabs that just have one
3124 * object freed. If they are used for allocation then they can be
3125 * filled up again with minimal effort. The slab will never hit the
3126 * per node partial lists and therefore no locking will be required.
3127 *
3128 * This setting also determines
3129 *
3130 * A) The number of objects from per cpu partial slabs dumped to the
3131 * per node list when we reach the limit.
Alex Shi9f264902011-09-01 11:32:18 +08003132 * B) The number of objects in cpu partial slabs to extract from the
Chen Gangd0e0ac92013-07-15 09:05:29 +08003133 * per node list when we run out of per cpu objects. We only fetch
3134 * 50% to keep some capacity around for frees.
Christoph Lameter49e22582011-08-09 16:12:27 -05003135 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09003136 if (!kmem_cache_has_cpu_partial(s))
Christoph Lameter8f1e33d2011-11-23 09:24:27 -06003137 s->cpu_partial = 0;
3138 else if (s->size >= PAGE_SIZE)
Christoph Lameter49e22582011-08-09 16:12:27 -05003139 s->cpu_partial = 2;
3140 else if (s->size >= 1024)
3141 s->cpu_partial = 6;
3142 else if (s->size >= 256)
3143 s->cpu_partial = 13;
3144 else
3145 s->cpu_partial = 30;
3146
Christoph Lameter81819f02007-05-06 14:49:36 -07003147#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05003148 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07003149#endif
Christoph Lameter55136592010-08-20 12:37:13 -05003150 if (!init_kmem_cache_nodes(s))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003151 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07003152
Christoph Lameter55136592010-08-20 12:37:13 -05003153 if (alloc_kmem_cache_cpus(s))
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003154 return 0;
Christoph Lameterff120592009-12-18 16:26:22 -06003155
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003156 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003157error:
3158 if (flags & SLAB_PANIC)
3159 panic("Cannot create slab %s size=%lu realsize=%u "
3160 "order=%u offset=%u flags=%lx\n",
Chen Gangd0e0ac92013-07-15 09:05:29 +08003161 s->name, (unsigned long)s->size, s->size,
3162 oo_order(s->oo), s->offset, flags);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003163 return -EINVAL;
Christoph Lameter81819f02007-05-06 14:49:36 -07003164}
Christoph Lameter81819f02007-05-06 14:49:36 -07003165
Christoph Lameter33b12c32008-04-25 12:22:43 -07003166static void list_slab_objects(struct kmem_cache *s, struct page *page,
3167 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07003168{
Christoph Lameter33b12c32008-04-25 12:22:43 -07003169#ifdef CONFIG_SLUB_DEBUG
3170 void *addr = page_address(page);
3171 void *p;
Namhyung Kima5dd5c12010-09-29 21:02:13 +09003172 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3173 sizeof(long), GFP_ATOMIC);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003174 if (!map)
3175 return;
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003176 slab_err(s, page, text, s->name);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003177 slab_lock(page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003178
Christoph Lameter5f80b132011-04-15 14:48:13 -05003179 get_map(s, page, map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003180 for_each_object(p, s, addr, page->objects) {
3181
3182 if (!test_bit(slab_index(p, s, addr), map)) {
3183 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3184 p, p - addr);
3185 print_tracking(s, p);
3186 }
3187 }
3188 slab_unlock(page);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003189 kfree(map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003190#endif
3191}
3192
Christoph Lameter81819f02007-05-06 14:49:36 -07003193/*
Christoph Lameter599870b2008-04-23 12:36:52 -07003194 * Attempt to free all partial slabs on a node.
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003195 * This is called from kmem_cache_close(). We must be the last thread
3196 * using the cache and therefore we do not need to lock anymore.
Christoph Lameter81819f02007-05-06 14:49:36 -07003197 */
Christoph Lameter599870b2008-04-23 12:36:52 -07003198static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003199{
Christoph Lameter81819f02007-05-06 14:49:36 -07003200 struct page *page, *h;
3201
Christoph Lameter33b12c32008-04-25 12:22:43 -07003202 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003203 if (!page->inuse) {
Steven Rostedt1e4dd942014-02-10 14:25:46 -08003204 __remove_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07003205 discard_slab(s, page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003206 } else {
3207 list_slab_objects(s, page,
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003208 "Objects remaining in %s on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07003209 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07003210 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003211}
3212
3213/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003214 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07003215 */
Christoph Lameter0c710012007-07-17 04:03:24 -07003216static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003217{
3218 int node;
3219
3220 flush_all(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003221 /* Attempt to free all objects */
Christoph Lameterf64dc582007-10-16 01:25:33 -07003222 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003223 struct kmem_cache_node *n = get_node(s, node);
3224
Christoph Lameter599870b2008-04-23 12:36:52 -07003225 free_partial(s, n);
3226 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07003227 return 1;
3228 }
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003229 free_percpu(s->cpu_slab);
Christoph Lameter81819f02007-05-06 14:49:36 -07003230 free_kmem_cache_nodes(s);
3231 return 0;
3232}
3233
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003234int __kmem_cache_shutdown(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003235{
Christoph Lameter12c36672012-09-04 23:38:33 +00003236 int rc = kmem_cache_close(s);
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003237
Glauber Costa5413dfb2012-12-18 14:23:13 -08003238 if (!rc) {
3239 /*
Vladimir Davydov421af242014-04-03 14:48:22 -07003240 * Since slab_attr_store may take the slab_mutex, we should
3241 * release the lock while removing the sysfs entry in order to
3242 * avoid a deadlock. Because this is pretty much the last
Glauber Costa5413dfb2012-12-18 14:23:13 -08003243 * operation we do and the lock will be released shortly after
3244 * that in slab_common.c, we could just move sysfs_slab_remove
3245 * to a later point in common code. We should do that when we
3246 * have a common sysfs framework for all allocators.
3247 */
3248 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003249 sysfs_slab_remove(s);
Glauber Costa5413dfb2012-12-18 14:23:13 -08003250 mutex_lock(&slab_mutex);
3251 }
Christoph Lameter12c36672012-09-04 23:38:33 +00003252
3253 return rc;
Christoph Lameter81819f02007-05-06 14:49:36 -07003254}
Christoph Lameter81819f02007-05-06 14:49:36 -07003255
3256/********************************************************************
3257 * Kmalloc subsystem
3258 *******************************************************************/
3259
Christoph Lameter81819f02007-05-06 14:49:36 -07003260static int __init setup_slub_min_order(char *str)
3261{
Pekka Enberg06428782008-01-07 23:20:27 -08003262 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003263
3264 return 1;
3265}
3266
3267__setup("slub_min_order=", setup_slub_min_order);
3268
3269static int __init setup_slub_max_order(char *str)
3270{
Pekka Enberg06428782008-01-07 23:20:27 -08003271 get_option(&str, &slub_max_order);
David Rientjes818cf592009-04-23 09:58:22 +03003272 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003273
3274 return 1;
3275}
3276
3277__setup("slub_max_order=", setup_slub_max_order);
3278
3279static int __init setup_slub_min_objects(char *str)
3280{
Pekka Enberg06428782008-01-07 23:20:27 -08003281 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07003282
3283 return 1;
3284}
3285
3286__setup("slub_min_objects=", setup_slub_min_objects);
3287
3288static int __init setup_slub_nomerge(char *str)
3289{
3290 slub_nomerge = 1;
3291 return 1;
3292}
3293
3294__setup("slub_nomerge", setup_slub_nomerge);
3295
Christoph Lameter81819f02007-05-06 14:49:36 -07003296void *__kmalloc(size_t size, gfp_t flags)
3297{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003298 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003299 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003300
Christoph Lameter95a05b42013-01-10 19:14:19 +00003301 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003302 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003303
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003304 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003305
3306 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003307 return s;
3308
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003309 ret = slab_alloc(s, flags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003310
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003311 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003312
3313 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003314}
3315EXPORT_SYMBOL(__kmalloc);
3316
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003317#ifdef CONFIG_NUMA
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003318static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3319{
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003320 struct page *page;
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003321 void *ptr = NULL;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003322
Glauber Costad79923f2012-12-18 14:22:48 -08003323 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003324 page = alloc_pages_node(node, flags, get_order(size));
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003325 if (page)
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003326 ptr = page_address(page);
3327
Roman Bobnievd56791b2013-10-08 15:58:57 -07003328 kmalloc_large_node_hook(ptr, size, flags);
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003329 return ptr;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003330}
3331
Christoph Lameter81819f02007-05-06 14:49:36 -07003332void *__kmalloc_node(size_t size, gfp_t flags, int node)
3333{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003334 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003335 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003336
Christoph Lameter95a05b42013-01-10 19:14:19 +00003337 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003338 ret = kmalloc_large_node(size, flags, node);
3339
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003340 trace_kmalloc_node(_RET_IP_, ret,
3341 size, PAGE_SIZE << get_order(size),
3342 flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003343
3344 return ret;
3345 }
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003346
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003347 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003348
3349 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003350 return s;
3351
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003352 ret = slab_alloc_node(s, flags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003353
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003354 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003355
3356 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003357}
3358EXPORT_SYMBOL(__kmalloc_node);
3359#endif
3360
3361size_t ksize(const void *object)
3362{
Christoph Lameter272c1d22007-06-08 13:46:49 -07003363 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07003364
Christoph Lameteref8b4522007-10-16 01:24:46 -07003365 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07003366 return 0;
3367
Vegard Nossum294a80a2007-12-04 23:45:30 -08003368 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08003369
Pekka Enberg76994412008-05-22 19:22:25 +03003370 if (unlikely(!PageSlab(page))) {
3371 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08003372 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03003373 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003374
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003375 return slab_ksize(page->slab_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003376}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02003377EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07003378
3379void kfree(const void *x)
3380{
Christoph Lameter81819f02007-05-06 14:49:36 -07003381 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08003382 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003383
Pekka Enberg2121db72009-03-25 11:05:57 +02003384 trace_kfree(_RET_IP_, x);
3385
Satyam Sharma2408c552007-10-16 01:24:44 -07003386 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07003387 return;
3388
Christoph Lameterb49af682007-05-06 14:49:41 -07003389 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003390 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07003391 BUG_ON(!PageCompound(page));
Roman Bobnievd56791b2013-10-08 15:58:57 -07003392 kfree_hook(x);
Glauber Costad79923f2012-12-18 14:22:48 -08003393 __free_memcg_kmem_pages(page, compound_order(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003394 return;
3395 }
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003396 slab_free(page->slab_cache, page, object, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07003397}
3398EXPORT_SYMBOL(kfree);
3399
Christoph Lameter2086d262007-05-06 14:49:46 -07003400/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003401 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3402 * the remaining slabs by the number of items in use. The slabs with the
3403 * most items in use come first. New allocations will then fill those up
3404 * and thus they can be removed from the partial lists.
3405 *
3406 * The slabs with the least items are placed last. This results in them
3407 * being allocated from last increasing the chance that the last objects
3408 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07003409 */
3410int kmem_cache_shrink(struct kmem_cache *s)
3411{
3412 int node;
3413 int i;
3414 struct kmem_cache_node *n;
3415 struct page *page;
3416 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03003417 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07003418 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03003419 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07003420 unsigned long flags;
3421
3422 if (!slabs_by_inuse)
3423 return -ENOMEM;
3424
3425 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003426 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07003427 n = get_node(s, node);
3428
3429 if (!n->nr_partial)
3430 continue;
3431
Christoph Lameter834f3d12008-04-14 19:11:31 +03003432 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07003433 INIT_LIST_HEAD(slabs_by_inuse + i);
3434
3435 spin_lock_irqsave(&n->list_lock, flags);
3436
3437 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003438 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07003439 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003440 * Note that concurrent frees may occur while we hold the
3441 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07003442 */
3443 list_for_each_entry_safe(page, t, &n->partial, lru) {
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003444 list_move(&page->lru, slabs_by_inuse + page->inuse);
3445 if (!page->inuse)
3446 n->nr_partial--;
Christoph Lameter2086d262007-05-06 14:49:46 -07003447 }
3448
Christoph Lameter2086d262007-05-06 14:49:46 -07003449 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003450 * Rebuild the partial list with the slabs filled up most
3451 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07003452 */
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003453 for (i = objects - 1; i > 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07003454 list_splice(slabs_by_inuse + i, n->partial.prev);
3455
Christoph Lameter2086d262007-05-06 14:49:46 -07003456 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003457
3458 /* Release empty slabs */
3459 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3460 discard_slab(s, page);
Christoph Lameter2086d262007-05-06 14:49:46 -07003461 }
3462
3463 kfree(slabs_by_inuse);
3464 return 0;
3465}
3466EXPORT_SYMBOL(kmem_cache_shrink);
3467
Yasunori Gotob9049e22007-10-21 16:41:37 -07003468static int slab_mem_going_offline_callback(void *arg)
3469{
3470 struct kmem_cache *s;
3471
Christoph Lameter18004c52012-07-06 15:25:12 -05003472 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003473 list_for_each_entry(s, &slab_caches, list)
3474 kmem_cache_shrink(s);
Christoph Lameter18004c52012-07-06 15:25:12 -05003475 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003476
3477 return 0;
3478}
3479
3480static void slab_mem_offline_callback(void *arg)
3481{
3482 struct kmem_cache_node *n;
3483 struct kmem_cache *s;
3484 struct memory_notify *marg = arg;
3485 int offline_node;
3486
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003487 offline_node = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003488
3489 /*
3490 * If the node still has available memory. we need kmem_cache_node
3491 * for it yet.
3492 */
3493 if (offline_node < 0)
3494 return;
3495
Christoph Lameter18004c52012-07-06 15:25:12 -05003496 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003497 list_for_each_entry(s, &slab_caches, list) {
3498 n = get_node(s, offline_node);
3499 if (n) {
3500 /*
3501 * if n->nr_slabs > 0, slabs still exist on the node
3502 * that is going down. We were unable to free them,
Adam Buchbinderc9404c92009-12-18 15:40:42 -05003503 * and offline_pages() function shouldn't call this
Yasunori Gotob9049e22007-10-21 16:41:37 -07003504 * callback. So, we must fail.
3505 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003506 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07003507
3508 s->node[offline_node] = NULL;
Christoph Lameter8de66a02010-08-25 14:51:14 -05003509 kmem_cache_free(kmem_cache_node, n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003510 }
3511 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003512 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003513}
3514
3515static int slab_mem_going_online_callback(void *arg)
3516{
3517 struct kmem_cache_node *n;
3518 struct kmem_cache *s;
3519 struct memory_notify *marg = arg;
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003520 int nid = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003521 int ret = 0;
3522
3523 /*
3524 * If the node's memory is already available, then kmem_cache_node is
3525 * already created. Nothing to do.
3526 */
3527 if (nid < 0)
3528 return 0;
3529
3530 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07003531 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07003532 * allocate a kmem_cache_node structure in order to bring the node
3533 * online.
3534 */
Christoph Lameter18004c52012-07-06 15:25:12 -05003535 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003536 list_for_each_entry(s, &slab_caches, list) {
3537 /*
3538 * XXX: kmem_cache_alloc_node will fallback to other nodes
3539 * since memory is not yet available from the node that
3540 * is brought up.
3541 */
Christoph Lameter8de66a02010-08-25 14:51:14 -05003542 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003543 if (!n) {
3544 ret = -ENOMEM;
3545 goto out;
3546 }
Joonsoo Kim40534972012-05-11 00:50:47 +09003547 init_kmem_cache_node(n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003548 s->node[nid] = n;
3549 }
3550out:
Christoph Lameter18004c52012-07-06 15:25:12 -05003551 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003552 return ret;
3553}
3554
3555static int slab_memory_callback(struct notifier_block *self,
3556 unsigned long action, void *arg)
3557{
3558 int ret = 0;
3559
3560 switch (action) {
3561 case MEM_GOING_ONLINE:
3562 ret = slab_mem_going_online_callback(arg);
3563 break;
3564 case MEM_GOING_OFFLINE:
3565 ret = slab_mem_going_offline_callback(arg);
3566 break;
3567 case MEM_OFFLINE:
3568 case MEM_CANCEL_ONLINE:
3569 slab_mem_offline_callback(arg);
3570 break;
3571 case MEM_ONLINE:
3572 case MEM_CANCEL_OFFLINE:
3573 break;
3574 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08003575 if (ret)
3576 ret = notifier_from_errno(ret);
3577 else
3578 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003579 return ret;
3580}
3581
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003582static struct notifier_block slab_memory_callback_nb = {
3583 .notifier_call = slab_memory_callback,
3584 .priority = SLAB_CALLBACK_PRI,
3585};
Yasunori Gotob9049e22007-10-21 16:41:37 -07003586
Christoph Lameter81819f02007-05-06 14:49:36 -07003587/********************************************************************
3588 * Basic setup of slabs
3589 *******************************************************************/
3590
Christoph Lameter51df1142010-08-20 12:37:15 -05003591/*
3592 * Used for early kmem_cache structures that were allocated using
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003593 * the page allocator. Allocate them properly then fix up the pointers
3594 * that may be pointing to the wrong kmem_cache structure.
Christoph Lameter51df1142010-08-20 12:37:15 -05003595 */
3596
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003597static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
Christoph Lameter51df1142010-08-20 12:37:15 -05003598{
3599 int node;
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003600 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
Christoph Lameter51df1142010-08-20 12:37:15 -05003601
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003602 memcpy(s, static_cache, kmem_cache->object_size);
Christoph Lameter51df1142010-08-20 12:37:15 -05003603
Glauber Costa7d557b32013-02-22 20:20:00 +04003604 /*
3605 * This runs very early, and only the boot processor is supposed to be
3606 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3607 * IPIs around.
3608 */
3609 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter51df1142010-08-20 12:37:15 -05003610 for_each_node_state(node, N_NORMAL_MEMORY) {
3611 struct kmem_cache_node *n = get_node(s, node);
3612 struct page *p;
3613
3614 if (n) {
3615 list_for_each_entry(p, &n->partial, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003616 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003617
Li Zefan607bf322011-04-12 15:22:26 +08003618#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter51df1142010-08-20 12:37:15 -05003619 list_for_each_entry(p, &n->full, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003620 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003621#endif
3622 }
3623 }
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003624 list_add(&s->list, &slab_caches);
3625 return s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003626}
3627
Christoph Lameter81819f02007-05-06 14:49:36 -07003628void __init kmem_cache_init(void)
3629{
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003630 static __initdata struct kmem_cache boot_kmem_cache,
3631 boot_kmem_cache_node;
Christoph Lameter51df1142010-08-20 12:37:15 -05003632
Stanislaw Gruszkafc8d8622012-01-10 15:07:32 -08003633 if (debug_guardpage_minorder())
3634 slub_max_order = 0;
3635
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003636 kmem_cache_node = &boot_kmem_cache_node;
3637 kmem_cache = &boot_kmem_cache;
Christoph Lameter51df1142010-08-20 12:37:15 -05003638
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003639 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3640 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003641
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003642 register_hotmemory_notifier(&slab_memory_callback_nb);
Christoph Lameter81819f02007-05-06 14:49:36 -07003643
3644 /* Able to allocate the per node structures */
3645 slab_state = PARTIAL;
3646
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003647 create_boot_cache(kmem_cache, "kmem_cache",
3648 offsetof(struct kmem_cache, node) +
3649 nr_node_ids * sizeof(struct kmem_cache_node *),
3650 SLAB_HWCACHE_ALIGN);
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003651
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003652 kmem_cache = bootstrap(&boot_kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003653
Christoph Lameter51df1142010-08-20 12:37:15 -05003654 /*
3655 * Allocate kmem_cache_node properly from the kmem_cache slab.
3656 * kmem_cache_node is separately allocated so no need to
3657 * update any list pointers.
3658 */
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003659 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
Christoph Lameter51df1142010-08-20 12:37:15 -05003660
3661 /* Now we can use the kmem_cache to allocate kmalloc slabs */
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003662 create_kmalloc_caches(0);
Christoph Lameter81819f02007-05-06 14:49:36 -07003663
3664#ifdef CONFIG_SMP
3665 register_cpu_notifier(&slab_notifier);
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003666#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003667
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003668 printk(KERN_INFO
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003669 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003670 " CPUs=%d, Nodes=%d\n",
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003671 cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003672 slub_min_order, slub_max_order, slub_min_objects,
3673 nr_cpu_ids, nr_node_ids);
3674}
3675
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003676void __init kmem_cache_init_late(void)
3677{
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003678}
3679
Christoph Lameter81819f02007-05-06 14:49:36 -07003680/*
3681 * Find a mergeable slab cache
3682 */
3683static int slab_unmergeable(struct kmem_cache *s)
3684{
3685 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3686 return 1;
3687
Vladimir Davydova44cb942014-04-07 15:39:23 -07003688 if (!is_root_cache(s))
3689 return 1;
3690
Christoph Lameterc59def92007-05-16 22:10:50 -07003691 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003692 return 1;
3693
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003694 /*
3695 * We may have set a slab to be unmergeable during bootstrap.
3696 */
3697 if (s->refcount < 0)
3698 return 1;
3699
Christoph Lameter81819f02007-05-06 14:49:36 -07003700 return 0;
3701}
3702
Vladimir Davydova44cb942014-04-07 15:39:23 -07003703static struct kmem_cache *find_mergeable(size_t size, size_t align,
3704 unsigned long flags, const char *name, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003705{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003706 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003707
3708 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3709 return NULL;
3710
Christoph Lameterc59def92007-05-16 22:10:50 -07003711 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003712 return NULL;
3713
3714 size = ALIGN(size, sizeof(void *));
3715 align = calculate_alignment(flags, align, size);
3716 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003717 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003718
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003719 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003720 if (slab_unmergeable(s))
3721 continue;
3722
3723 if (size > s->size)
3724 continue;
3725
Christoph Lameterba0268a2007-09-11 15:24:11 -07003726 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Vladimir Davydova44cb942014-04-07 15:39:23 -07003727 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003728 /*
3729 * Check if alignment is compatible.
3730 * Courtesy of Adrian Drzewiecki
3731 */
Pekka Enberg06428782008-01-07 23:20:27 -08003732 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003733 continue;
3734
3735 if (s->size - size >= sizeof(void *))
3736 continue;
3737
3738 return s;
3739 }
3740 return NULL;
3741}
3742
Glauber Costa2633d7a2012-12-18 14:22:34 -08003743struct kmem_cache *
Vladimir Davydova44cb942014-04-07 15:39:23 -07003744__kmem_cache_alias(const char *name, size_t size, size_t align,
3745 unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003746{
3747 struct kmem_cache *s;
3748
Vladimir Davydova44cb942014-04-07 15:39:23 -07003749 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003750 if (s) {
Vladimir Davydov84d0ddd2014-04-07 15:39:29 -07003751 int i;
3752 struct kmem_cache *c;
3753
Christoph Lameter81819f02007-05-06 14:49:36 -07003754 s->refcount++;
Vladimir Davydov84d0ddd2014-04-07 15:39:29 -07003755
Christoph Lameter81819f02007-05-06 14:49:36 -07003756 /*
3757 * Adjust the object sizes so that we clear
3758 * the complete object on kzalloc.
3759 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003760 s->object_size = max(s->object_size, (int)size);
Christoph Lameter81819f02007-05-06 14:49:36 -07003761 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lameter6446faa2008-02-15 23:45:26 -08003762
Vladimir Davydov84d0ddd2014-04-07 15:39:29 -07003763 for_each_memcg_cache_index(i) {
3764 c = cache_from_memcg_idx(s, i);
3765 if (!c)
3766 continue;
3767 c->object_size = s->object_size;
3768 c->inuse = max_t(int, c->inuse,
3769 ALIGN(size, sizeof(void *)));
3770 }
3771
David Rientjes7b8f3b62008-12-17 22:09:46 -08003772 if (sysfs_slab_alias(s, name)) {
David Rientjes7b8f3b62008-12-17 22:09:46 -08003773 s->refcount--;
Christoph Lametercbb79692012-09-05 00:18:32 +00003774 s = NULL;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003775 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003776 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003777
Christoph Lametercbb79692012-09-05 00:18:32 +00003778 return s;
3779}
Pekka Enberg84c1cf62010-09-14 23:21:12 +03003780
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003781int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
Christoph Lametercbb79692012-09-05 00:18:32 +00003782{
Pekka Enbergaac3a162012-09-05 12:07:44 +03003783 int err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003784
Pekka Enbergaac3a162012-09-05 12:07:44 +03003785 err = kmem_cache_open(s, flags);
3786 if (err)
3787 return err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003788
Christoph Lameter45530c42012-11-28 16:23:07 +00003789 /* Mutex is not taken during early boot */
3790 if (slab_state <= UP)
3791 return 0;
3792
Glauber Costa107dab52012-12-18 14:23:05 -08003793 memcg_propagate_slab_attrs(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003794 err = sysfs_slab_add(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003795 if (err)
3796 kmem_cache_close(s);
3797
3798 return err;
Christoph Lameter81819f02007-05-06 14:49:36 -07003799}
Christoph Lameter81819f02007-05-06 14:49:36 -07003800
Christoph Lameter81819f02007-05-06 14:49:36 -07003801#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003802/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003803 * Use the cpu notifier to insure that the cpu slabs are flushed when
3804 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003805 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04003806static int slab_cpuup_callback(struct notifier_block *nfb,
Christoph Lameter81819f02007-05-06 14:49:36 -07003807 unsigned long action, void *hcpu)
3808{
3809 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003810 struct kmem_cache *s;
3811 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003812
3813 switch (action) {
3814 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003815 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003816 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003817 case CPU_DEAD_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05003818 mutex_lock(&slab_mutex);
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003819 list_for_each_entry(s, &slab_caches, list) {
3820 local_irq_save(flags);
3821 __flush_cpu_slab(s, cpu);
3822 local_irq_restore(flags);
3823 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003824 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003825 break;
3826 default:
3827 break;
3828 }
3829 return NOTIFY_OK;
3830}
3831
Paul Gortmaker0db06282013-06-19 14:53:51 -04003832static struct notifier_block slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003833 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003834};
Christoph Lameter81819f02007-05-06 14:49:36 -07003835
3836#endif
3837
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003838void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003839{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003840 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003841 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003842
Christoph Lameter95a05b42013-01-10 19:14:19 +00003843 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003844 return kmalloc_large(size, gfpflags);
3845
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003846 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003847
Satyam Sharma2408c552007-10-16 01:24:44 -07003848 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003849 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003850
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003851 ret = slab_alloc(s, gfpflags, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003852
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003853 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003854 trace_kmalloc(caller, ret, size, s->size, gfpflags);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003855
3856 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003857}
3858
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003859#ifdef CONFIG_NUMA
Christoph Lameter81819f02007-05-06 14:49:36 -07003860void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003861 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003862{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003863 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003864 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003865
Christoph Lameter95a05b42013-01-10 19:14:19 +00003866 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Xiaotian Fengd3e14aa2010-04-08 17:26:44 +08003867 ret = kmalloc_large_node(size, gfpflags, node);
3868
3869 trace_kmalloc_node(caller, ret,
3870 size, PAGE_SIZE << get_order(size),
3871 gfpflags, node);
3872
3873 return ret;
3874 }
Pekka Enbergeada35e2008-02-11 22:47:46 +02003875
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003876 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003877
Satyam Sharma2408c552007-10-16 01:24:44 -07003878 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003879 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003880
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003881 ret = slab_alloc_node(s, gfpflags, node, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003882
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003883 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003884 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003885
3886 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003887}
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003888#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003889
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003890#ifdef CONFIG_SYSFS
Christoph Lameter205ab992008-04-14 19:11:40 +03003891static int count_inuse(struct page *page)
3892{
3893 return page->inuse;
3894}
3895
3896static int count_total(struct page *page)
3897{
3898 return page->objects;
3899}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003900#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03003901
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003902#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter434e2452007-07-17 04:03:30 -07003903static int validate_slab(struct kmem_cache *s, struct page *page,
3904 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003905{
3906 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003907 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003908
3909 if (!check_slab(s, page) ||
3910 !on_freelist(s, page, NULL))
3911 return 0;
3912
3913 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003914 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003915
Christoph Lameter5f80b132011-04-15 14:48:13 -05003916 get_map(s, page, map);
3917 for_each_object(p, s, addr, page->objects) {
3918 if (test_bit(slab_index(p, s, addr), map))
3919 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3920 return 0;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003921 }
3922
Christoph Lameter224a88b2008-04-14 19:11:31 +03003923 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003924 if (!test_bit(slab_index(p, s, addr), map))
Tero Roponen37d57442010-12-01 20:04:20 +02003925 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003926 return 0;
3927 return 1;
3928}
3929
Christoph Lameter434e2452007-07-17 04:03:30 -07003930static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3931 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003932{
Christoph Lameter881db7f2011-06-01 12:25:53 -05003933 slab_lock(page);
3934 validate_slab(s, page, map);
3935 slab_unlock(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003936}
3937
Christoph Lameter434e2452007-07-17 04:03:30 -07003938static int validate_slab_node(struct kmem_cache *s,
3939 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003940{
3941 unsigned long count = 0;
3942 struct page *page;
3943 unsigned long flags;
3944
3945 spin_lock_irqsave(&n->list_lock, flags);
3946
3947 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003948 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003949 count++;
3950 }
3951 if (count != n->nr_partial)
3952 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3953 "counter=%ld\n", s->name, count, n->nr_partial);
3954
3955 if (!(s->flags & SLAB_STORE_USER))
3956 goto out;
3957
3958 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003959 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003960 count++;
3961 }
3962 if (count != atomic_long_read(&n->nr_slabs))
3963 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3964 "counter=%ld\n", s->name, count,
3965 atomic_long_read(&n->nr_slabs));
3966
3967out:
3968 spin_unlock_irqrestore(&n->list_lock, flags);
3969 return count;
3970}
3971
Christoph Lameter434e2452007-07-17 04:03:30 -07003972static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003973{
3974 int node;
3975 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003976 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003977 sizeof(unsigned long), GFP_KERNEL);
3978
3979 if (!map)
3980 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003981
3982 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003983 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003984 struct kmem_cache_node *n = get_node(s, node);
3985
Christoph Lameter434e2452007-07-17 04:03:30 -07003986 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003987 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003988 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003989 return count;
3990}
Christoph Lameter88a420e2007-05-06 14:49:45 -07003991/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003992 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003993 * and freed.
3994 */
3995
3996struct location {
3997 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003998 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003999 long long sum_time;
4000 long min_time;
4001 long max_time;
4002 long min_pid;
4003 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304004 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004005 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004006};
4007
4008struct loc_track {
4009 unsigned long max;
4010 unsigned long count;
4011 struct location *loc;
4012};
4013
4014static void free_loc_track(struct loc_track *t)
4015{
4016 if (t->max)
4017 free_pages((unsigned long)t->loc,
4018 get_order(sizeof(struct location) * t->max));
4019}
4020
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004021static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004022{
4023 struct location *l;
4024 int order;
4025
Christoph Lameter88a420e2007-05-06 14:49:45 -07004026 order = get_order(sizeof(struct location) * max);
4027
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004028 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004029 if (!l)
4030 return 0;
4031
4032 if (t->count) {
4033 memcpy(l, t->loc, sizeof(struct location) * t->count);
4034 free_loc_track(t);
4035 }
4036 t->max = max;
4037 t->loc = l;
4038 return 1;
4039}
4040
4041static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07004042 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004043{
4044 long start, end, pos;
4045 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004046 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004047 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004048
4049 start = -1;
4050 end = t->count;
4051
4052 for ( ; ; ) {
4053 pos = start + (end - start + 1) / 2;
4054
4055 /*
4056 * There is nothing at "end". If we end up there
4057 * we need to add something to before end.
4058 */
4059 if (pos == end)
4060 break;
4061
4062 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004063 if (track->addr == caddr) {
4064
4065 l = &t->loc[pos];
4066 l->count++;
4067 if (track->when) {
4068 l->sum_time += age;
4069 if (age < l->min_time)
4070 l->min_time = age;
4071 if (age > l->max_time)
4072 l->max_time = age;
4073
4074 if (track->pid < l->min_pid)
4075 l->min_pid = track->pid;
4076 if (track->pid > l->max_pid)
4077 l->max_pid = track->pid;
4078
Rusty Russell174596a2009-01-01 10:12:29 +10304079 cpumask_set_cpu(track->cpu,
4080 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004081 }
4082 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004083 return 1;
4084 }
4085
Christoph Lameter45edfa52007-05-09 02:32:45 -07004086 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004087 end = pos;
4088 else
4089 start = pos;
4090 }
4091
4092 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07004093 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07004094 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004095 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004096 return 0;
4097
4098 l = t->loc + pos;
4099 if (pos < t->count)
4100 memmove(l + 1, l,
4101 (t->count - pos) * sizeof(struct location));
4102 t->count++;
4103 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004104 l->addr = track->addr;
4105 l->sum_time = age;
4106 l->min_time = age;
4107 l->max_time = age;
4108 l->min_pid = track->pid;
4109 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304110 cpumask_clear(to_cpumask(l->cpus));
4111 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004112 nodes_clear(l->nodes);
4113 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004114 return 1;
4115}
4116
4117static void process_slab(struct loc_track *t, struct kmem_cache *s,
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004118 struct page *page, enum track_item alloc,
Namhyung Kima5dd5c12010-09-29 21:02:13 +09004119 unsigned long *map)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004120{
Christoph Lametera973e9d2008-03-01 13:40:44 -08004121 void *addr = page_address(page);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004122 void *p;
4123
Christoph Lameter39b26462008-04-14 19:11:30 +03004124 bitmap_zero(map, page->objects);
Christoph Lameter5f80b132011-04-15 14:48:13 -05004125 get_map(s, page, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004126
Christoph Lameter224a88b2008-04-14 19:11:31 +03004127 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07004128 if (!test_bit(slab_index(p, s, addr), map))
4129 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07004130}
4131
4132static int list_locations(struct kmem_cache *s, char *buf,
4133 enum track_item alloc)
4134{
Harvey Harrisone374d482008-01-31 15:20:50 -08004135 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004136 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004137 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07004138 int node;
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004139 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4140 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004141
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004142 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4143 GFP_TEMPORARY)) {
4144 kfree(map);
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004145 return sprintf(buf, "Out of memory\n");
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004146 }
Christoph Lameter88a420e2007-05-06 14:49:45 -07004147 /* Push back cpu slabs */
4148 flush_all(s);
4149
Christoph Lameterf64dc582007-10-16 01:25:33 -07004150 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07004151 struct kmem_cache_node *n = get_node(s, node);
4152 unsigned long flags;
4153 struct page *page;
4154
Christoph Lameter9e869432007-08-22 14:01:56 -07004155 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004156 continue;
4157
4158 spin_lock_irqsave(&n->list_lock, flags);
4159 list_for_each_entry(page, &n->partial, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004160 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004161 list_for_each_entry(page, &n->full, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004162 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004163 spin_unlock_irqrestore(&n->list_lock, flags);
4164 }
4165
4166 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07004167 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07004168
Hugh Dickins9c246242008-12-09 13:14:27 -08004169 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004170 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08004171 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004172
4173 if (l->addr)
Joe Perches62c70bc2011-01-13 15:45:52 -08004174 len += sprintf(buf + len, "%pS", (void *)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004175 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004176 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07004177
4178 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08004179 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07004180 l->min_time,
4181 (long)div_u64(l->sum_time, l->count),
4182 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004183 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08004184 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004185 l->min_time);
4186
4187 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08004188 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004189 l->min_pid, l->max_pid);
4190 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004191 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004192 l->min_pid);
4193
Rusty Russell174596a2009-01-01 10:12:29 +10304194 if (num_online_cpus() > 1 &&
4195 !cpumask_empty(to_cpumask(l->cpus)) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004196 len < PAGE_SIZE - 60) {
4197 len += sprintf(buf + len, " cpus=");
Chen Gangd0e0ac92013-07-15 09:05:29 +08004198 len += cpulist_scnprintf(buf + len,
4199 PAGE_SIZE - len - 50,
Rusty Russell174596a2009-01-01 10:12:29 +10304200 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004201 }
4202
Christoph Lameter62bc62a2009-06-16 15:32:15 -07004203 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004204 len < PAGE_SIZE - 60) {
4205 len += sprintf(buf + len, " nodes=");
Chen Gangd0e0ac92013-07-15 09:05:29 +08004206 len += nodelist_scnprintf(buf + len,
4207 PAGE_SIZE - len - 50,
4208 l->nodes);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004209 }
4210
Harvey Harrisone374d482008-01-31 15:20:50 -08004211 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07004212 }
4213
4214 free_loc_track(&t);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004215 kfree(map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004216 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08004217 len += sprintf(buf, "No data\n");
4218 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004219}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004220#endif
Christoph Lameter88a420e2007-05-06 14:49:45 -07004221
Christoph Lametera5a84752010-10-05 13:57:27 -05004222#ifdef SLUB_RESILIENCY_TEST
4223static void resiliency_test(void)
4224{
4225 u8 *p;
4226
Christoph Lameter95a05b42013-01-10 19:14:19 +00004227 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
Christoph Lametera5a84752010-10-05 13:57:27 -05004228
4229 printk(KERN_ERR "SLUB resiliency testing\n");
4230 printk(KERN_ERR "-----------------------\n");
4231 printk(KERN_ERR "A. Corruption after allocation\n");
4232
4233 p = kzalloc(16, GFP_KERNEL);
4234 p[16] = 0x12;
4235 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4236 " 0x12->0x%p\n\n", p + 16);
4237
4238 validate_slab_cache(kmalloc_caches[4]);
4239
4240 /* Hmmm... The next two are dangerous */
4241 p = kzalloc(32, GFP_KERNEL);
4242 p[32 + sizeof(void *)] = 0x34;
4243 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4244 " 0x34 -> -0x%p\n", p);
4245 printk(KERN_ERR
4246 "If allocated object is overwritten then not detectable\n\n");
4247
4248 validate_slab_cache(kmalloc_caches[5]);
4249 p = kzalloc(64, GFP_KERNEL);
4250 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4251 *p = 0x56;
4252 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4253 p);
4254 printk(KERN_ERR
4255 "If allocated object is overwritten then not detectable\n\n");
4256 validate_slab_cache(kmalloc_caches[6]);
4257
4258 printk(KERN_ERR "\nB. Corruption after free\n");
4259 p = kzalloc(128, GFP_KERNEL);
4260 kfree(p);
4261 *p = 0x78;
4262 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4263 validate_slab_cache(kmalloc_caches[7]);
4264
4265 p = kzalloc(256, GFP_KERNEL);
4266 kfree(p);
4267 p[50] = 0x9a;
4268 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4269 p);
4270 validate_slab_cache(kmalloc_caches[8]);
4271
4272 p = kzalloc(512, GFP_KERNEL);
4273 kfree(p);
4274 p[512] = 0xab;
4275 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4276 validate_slab_cache(kmalloc_caches[9]);
4277}
4278#else
4279#ifdef CONFIG_SYSFS
4280static void resiliency_test(void) {};
4281#endif
4282#endif
4283
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004284#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -07004285enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03004286 SL_ALL, /* All slabs */
4287 SL_PARTIAL, /* Only partially allocated slabs */
4288 SL_CPU, /* Only slabs used for cpu caches */
4289 SL_OBJECTS, /* Determine allocated objects not slabs */
4290 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07004291};
4292
Christoph Lameter205ab992008-04-14 19:11:40 +03004293#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004294#define SO_PARTIAL (1 << SL_PARTIAL)
4295#define SO_CPU (1 << SL_CPU)
4296#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03004297#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004298
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004299static ssize_t show_slab_objects(struct kmem_cache *s,
4300 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07004301{
4302 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07004303 int node;
4304 int x;
4305 unsigned long *nodes;
Christoph Lameter81819f02007-05-06 14:49:36 -07004306
Chen Gange35e1a92013-07-12 08:23:48 +08004307 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004308 if (!nodes)
4309 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07004310
Christoph Lameter205ab992008-04-14 19:11:40 +03004311 if (flags & SO_CPU) {
4312 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07004313
Christoph Lameter205ab992008-04-14 19:11:40 +03004314 for_each_possible_cpu(cpu) {
Chen Gangd0e0ac92013-07-15 09:05:29 +08004315 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4316 cpu);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004317 int node;
Christoph Lameter49e22582011-08-09 16:12:27 -05004318 struct page *page;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004319
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004320 page = ACCESS_ONCE(c->page);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004321 if (!page)
4322 continue;
Christoph Lameter205ab992008-04-14 19:11:40 +03004323
Christoph Lameterec3ab082012-05-09 10:09:56 -05004324 node = page_to_nid(page);
4325 if (flags & SO_TOTAL)
4326 x = page->objects;
4327 else if (flags & SO_OBJECTS)
4328 x = page->inuse;
4329 else
4330 x = 1;
Christoph Lameter49e22582011-08-09 16:12:27 -05004331
Christoph Lameterec3ab082012-05-09 10:09:56 -05004332 total += x;
4333 nodes[node] += x;
4334
4335 page = ACCESS_ONCE(c->partial);
Christoph Lameter49e22582011-08-09 16:12:27 -05004336 if (page) {
Li Zefan8afb1472013-09-10 11:43:37 +08004337 node = page_to_nid(page);
4338 if (flags & SO_TOTAL)
4339 WARN_ON_ONCE(1);
4340 else if (flags & SO_OBJECTS)
4341 WARN_ON_ONCE(1);
4342 else
4343 x = page->pages;
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004344 total += x;
4345 nodes[node] += x;
Christoph Lameter49e22582011-08-09 16:12:27 -05004346 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004347 }
4348 }
4349
Christoph Lameter04d94872011-01-10 10:15:15 -06004350 lock_memory_hotplug();
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004351#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03004352 if (flags & SO_ALL) {
4353 for_each_node_state(node, N_NORMAL_MEMORY) {
4354 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07004355
Chen Gangd0e0ac92013-07-15 09:05:29 +08004356 if (flags & SO_TOTAL)
4357 x = atomic_long_read(&n->total_objects);
4358 else if (flags & SO_OBJECTS)
4359 x = atomic_long_read(&n->total_objects) -
4360 count_partial(n, count_free);
Christoph Lameter205ab992008-04-14 19:11:40 +03004361 else
4362 x = atomic_long_read(&n->nr_slabs);
4363 total += x;
4364 nodes[node] += x;
4365 }
4366
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004367 } else
4368#endif
4369 if (flags & SO_PARTIAL) {
Christoph Lameter205ab992008-04-14 19:11:40 +03004370 for_each_node_state(node, N_NORMAL_MEMORY) {
4371 struct kmem_cache_node *n = get_node(s, node);
4372
4373 if (flags & SO_TOTAL)
4374 x = count_partial(n, count_total);
4375 else if (flags & SO_OBJECTS)
4376 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07004377 else
4378 x = n->nr_partial;
4379 total += x;
4380 nodes[node] += x;
4381 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004382 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004383 x = sprintf(buf, "%lu", total);
4384#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07004385 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07004386 if (nodes[node])
4387 x += sprintf(buf + x, " N%d=%lu",
4388 node, nodes[node]);
4389#endif
Christoph Lameter04d94872011-01-10 10:15:15 -06004390 unlock_memory_hotplug();
Christoph Lameter81819f02007-05-06 14:49:36 -07004391 kfree(nodes);
4392 return x + sprintf(buf + x, "\n");
4393}
4394
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004395#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07004396static int any_slab_objects(struct kmem_cache *s)
4397{
4398 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004399
4400 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004401 struct kmem_cache_node *n = get_node(s, node);
4402
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004403 if (!n)
4404 continue;
4405
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07004406 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07004407 return 1;
4408 }
4409 return 0;
4410}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004411#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004412
4413#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
Phil Carmody497888c2011-07-14 15:07:13 +03004414#define to_slab(n) container_of(n, struct kmem_cache, kobj)
Christoph Lameter81819f02007-05-06 14:49:36 -07004415
4416struct slab_attribute {
4417 struct attribute attr;
4418 ssize_t (*show)(struct kmem_cache *s, char *buf);
4419 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4420};
4421
4422#define SLAB_ATTR_RO(_name) \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004423 static struct slab_attribute _name##_attr = \
4424 __ATTR(_name, 0400, _name##_show, NULL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004425
4426#define SLAB_ATTR(_name) \
4427 static struct slab_attribute _name##_attr = \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004428 __ATTR(_name, 0600, _name##_show, _name##_store)
Christoph Lameter81819f02007-05-06 14:49:36 -07004429
Christoph Lameter81819f02007-05-06 14:49:36 -07004430static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4431{
4432 return sprintf(buf, "%d\n", s->size);
4433}
4434SLAB_ATTR_RO(slab_size);
4435
4436static ssize_t align_show(struct kmem_cache *s, char *buf)
4437{
4438 return sprintf(buf, "%d\n", s->align);
4439}
4440SLAB_ATTR_RO(align);
4441
4442static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4443{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004444 return sprintf(buf, "%d\n", s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -07004445}
4446SLAB_ATTR_RO(object_size);
4447
4448static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4449{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004450 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004451}
4452SLAB_ATTR_RO(objs_per_slab);
4453
Christoph Lameter06b285d2008-04-14 19:11:41 +03004454static ssize_t order_store(struct kmem_cache *s,
4455 const char *buf, size_t length)
4456{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004457 unsigned long order;
4458 int err;
4459
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004460 err = kstrtoul(buf, 10, &order);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004461 if (err)
4462 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004463
4464 if (order > slub_max_order || order < slub_min_order)
4465 return -EINVAL;
4466
4467 calculate_sizes(s, order);
4468 return length;
4469}
4470
Christoph Lameter81819f02007-05-06 14:49:36 -07004471static ssize_t order_show(struct kmem_cache *s, char *buf)
4472{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004473 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004474}
Christoph Lameter06b285d2008-04-14 19:11:41 +03004475SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07004476
David Rientjes73d342b2009-02-22 17:40:09 -08004477static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4478{
4479 return sprintf(buf, "%lu\n", s->min_partial);
4480}
4481
4482static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4483 size_t length)
4484{
4485 unsigned long min;
4486 int err;
4487
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004488 err = kstrtoul(buf, 10, &min);
David Rientjes73d342b2009-02-22 17:40:09 -08004489 if (err)
4490 return err;
4491
David Rientjesc0bdb232009-02-25 09:16:35 +02004492 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08004493 return length;
4494}
4495SLAB_ATTR(min_partial);
4496
Christoph Lameter49e22582011-08-09 16:12:27 -05004497static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4498{
4499 return sprintf(buf, "%u\n", s->cpu_partial);
4500}
4501
4502static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4503 size_t length)
4504{
4505 unsigned long objects;
4506 int err;
4507
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004508 err = kstrtoul(buf, 10, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05004509 if (err)
4510 return err;
Joonsoo Kim345c9052013-06-19 14:05:52 +09004511 if (objects && !kmem_cache_has_cpu_partial(s))
David Rientjes74ee4ef2012-01-09 13:19:45 -08004512 return -EINVAL;
Christoph Lameter49e22582011-08-09 16:12:27 -05004513
4514 s->cpu_partial = objects;
4515 flush_all(s);
4516 return length;
4517}
4518SLAB_ATTR(cpu_partial);
4519
Christoph Lameter81819f02007-05-06 14:49:36 -07004520static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4521{
Joe Perches62c70bc2011-01-13 15:45:52 -08004522 if (!s->ctor)
4523 return 0;
4524 return sprintf(buf, "%pS\n", s->ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07004525}
4526SLAB_ATTR_RO(ctor);
4527
Christoph Lameter81819f02007-05-06 14:49:36 -07004528static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4529{
4530 return sprintf(buf, "%d\n", s->refcount - 1);
4531}
4532SLAB_ATTR_RO(aliases);
4533
Christoph Lameter81819f02007-05-06 14:49:36 -07004534static ssize_t partial_show(struct kmem_cache *s, char *buf)
4535{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004536 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004537}
4538SLAB_ATTR_RO(partial);
4539
4540static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4541{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004542 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07004543}
4544SLAB_ATTR_RO(cpu_slabs);
4545
4546static ssize_t objects_show(struct kmem_cache *s, char *buf)
4547{
Christoph Lameter205ab992008-04-14 19:11:40 +03004548 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07004549}
4550SLAB_ATTR_RO(objects);
4551
Christoph Lameter205ab992008-04-14 19:11:40 +03004552static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4553{
4554 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4555}
4556SLAB_ATTR_RO(objects_partial);
4557
Christoph Lameter49e22582011-08-09 16:12:27 -05004558static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4559{
4560 int objects = 0;
4561 int pages = 0;
4562 int cpu;
4563 int len;
4564
4565 for_each_online_cpu(cpu) {
4566 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4567
4568 if (page) {
4569 pages += page->pages;
4570 objects += page->pobjects;
4571 }
4572 }
4573
4574 len = sprintf(buf, "%d(%d)", objects, pages);
4575
4576#ifdef CONFIG_SMP
4577 for_each_online_cpu(cpu) {
4578 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4579
4580 if (page && len < PAGE_SIZE - 20)
4581 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4582 page->pobjects, page->pages);
4583 }
4584#endif
4585 return len + sprintf(buf + len, "\n");
4586}
4587SLAB_ATTR_RO(slabs_cpu_partial);
4588
Christoph Lameter81819f02007-05-06 14:49:36 -07004589static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4590{
4591 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4592}
4593
4594static ssize_t reclaim_account_store(struct kmem_cache *s,
4595 const char *buf, size_t length)
4596{
4597 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4598 if (buf[0] == '1')
4599 s->flags |= SLAB_RECLAIM_ACCOUNT;
4600 return length;
4601}
4602SLAB_ATTR(reclaim_account);
4603
4604static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4605{
Christoph Lameter5af60832007-05-06 14:49:56 -07004606 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07004607}
4608SLAB_ATTR_RO(hwcache_align);
4609
4610#ifdef CONFIG_ZONE_DMA
4611static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4612{
4613 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4614}
4615SLAB_ATTR_RO(cache_dma);
4616#endif
4617
4618static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4619{
4620 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4621}
4622SLAB_ATTR_RO(destroy_by_rcu);
4623
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004624static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4625{
4626 return sprintf(buf, "%d\n", s->reserved);
4627}
4628SLAB_ATTR_RO(reserved);
4629
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004630#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004631static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4632{
4633 return show_slab_objects(s, buf, SO_ALL);
4634}
4635SLAB_ATTR_RO(slabs);
4636
4637static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4638{
4639 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4640}
4641SLAB_ATTR_RO(total_objects);
4642
4643static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4644{
4645 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4646}
4647
4648static ssize_t sanity_checks_store(struct kmem_cache *s,
4649 const char *buf, size_t length)
4650{
4651 s->flags &= ~SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004652 if (buf[0] == '1') {
4653 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004654 s->flags |= SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004655 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004656 return length;
4657}
4658SLAB_ATTR(sanity_checks);
4659
4660static ssize_t trace_show(struct kmem_cache *s, char *buf)
4661{
4662 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4663}
4664
4665static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4666 size_t length)
4667{
4668 s->flags &= ~SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004669 if (buf[0] == '1') {
4670 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004671 s->flags |= SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004672 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004673 return length;
4674}
4675SLAB_ATTR(trace);
4676
Christoph Lameter81819f02007-05-06 14:49:36 -07004677static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4678{
4679 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4680}
4681
4682static ssize_t red_zone_store(struct kmem_cache *s,
4683 const char *buf, size_t length)
4684{
4685 if (any_slab_objects(s))
4686 return -EBUSY;
4687
4688 s->flags &= ~SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004689 if (buf[0] == '1') {
4690 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004691 s->flags |= SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004692 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004693 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004694 return length;
4695}
4696SLAB_ATTR(red_zone);
4697
4698static ssize_t poison_show(struct kmem_cache *s, char *buf)
4699{
4700 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4701}
4702
4703static ssize_t poison_store(struct kmem_cache *s,
4704 const char *buf, size_t length)
4705{
4706 if (any_slab_objects(s))
4707 return -EBUSY;
4708
4709 s->flags &= ~SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004710 if (buf[0] == '1') {
4711 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004712 s->flags |= SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004713 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004714 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004715 return length;
4716}
4717SLAB_ATTR(poison);
4718
4719static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4720{
4721 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4722}
4723
4724static ssize_t store_user_store(struct kmem_cache *s,
4725 const char *buf, size_t length)
4726{
4727 if (any_slab_objects(s))
4728 return -EBUSY;
4729
4730 s->flags &= ~SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004731 if (buf[0] == '1') {
4732 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004733 s->flags |= SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004734 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004735 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004736 return length;
4737}
4738SLAB_ATTR(store_user);
4739
Christoph Lameter53e15af2007-05-06 14:49:43 -07004740static ssize_t validate_show(struct kmem_cache *s, char *buf)
4741{
4742 return 0;
4743}
4744
4745static ssize_t validate_store(struct kmem_cache *s,
4746 const char *buf, size_t length)
4747{
Christoph Lameter434e2452007-07-17 04:03:30 -07004748 int ret = -EINVAL;
4749
4750 if (buf[0] == '1') {
4751 ret = validate_slab_cache(s);
4752 if (ret >= 0)
4753 ret = length;
4754 }
4755 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004756}
4757SLAB_ATTR(validate);
Christoph Lametera5a84752010-10-05 13:57:27 -05004758
4759static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4760{
4761 if (!(s->flags & SLAB_STORE_USER))
4762 return -ENOSYS;
4763 return list_locations(s, buf, TRACK_ALLOC);
4764}
4765SLAB_ATTR_RO(alloc_calls);
4766
4767static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4768{
4769 if (!(s->flags & SLAB_STORE_USER))
4770 return -ENOSYS;
4771 return list_locations(s, buf, TRACK_FREE);
4772}
4773SLAB_ATTR_RO(free_calls);
4774#endif /* CONFIG_SLUB_DEBUG */
4775
4776#ifdef CONFIG_FAILSLAB
4777static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4778{
4779 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4780}
4781
4782static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4783 size_t length)
4784{
4785 s->flags &= ~SLAB_FAILSLAB;
4786 if (buf[0] == '1')
4787 s->flags |= SLAB_FAILSLAB;
4788 return length;
4789}
4790SLAB_ATTR(failslab);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004791#endif
Christoph Lameter53e15af2007-05-06 14:49:43 -07004792
Christoph Lameter2086d262007-05-06 14:49:46 -07004793static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4794{
4795 return 0;
4796}
4797
4798static ssize_t shrink_store(struct kmem_cache *s,
4799 const char *buf, size_t length)
4800{
4801 if (buf[0] == '1') {
4802 int rc = kmem_cache_shrink(s);
4803
4804 if (rc)
4805 return rc;
4806 } else
4807 return -EINVAL;
4808 return length;
4809}
4810SLAB_ATTR(shrink);
4811
Christoph Lameter81819f02007-05-06 14:49:36 -07004812#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004813static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004814{
Christoph Lameter98246012008-01-07 23:20:26 -08004815 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004816}
4817
Christoph Lameter98246012008-01-07 23:20:26 -08004818static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004819 const char *buf, size_t length)
4820{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004821 unsigned long ratio;
4822 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004823
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004824 err = kstrtoul(buf, 10, &ratio);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004825 if (err)
4826 return err;
4827
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004828 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004829 s->remote_node_defrag_ratio = ratio * 10;
4830
Christoph Lameter81819f02007-05-06 14:49:36 -07004831 return length;
4832}
Christoph Lameter98246012008-01-07 23:20:26 -08004833SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004834#endif
4835
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004836#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004837static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4838{
4839 unsigned long sum = 0;
4840 int cpu;
4841 int len;
4842 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4843
4844 if (!data)
4845 return -ENOMEM;
4846
4847 for_each_online_cpu(cpu) {
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004848 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004849
4850 data[cpu] = x;
4851 sum += x;
4852 }
4853
4854 len = sprintf(buf, "%lu", sum);
4855
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004856#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004857 for_each_online_cpu(cpu) {
4858 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004859 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004860 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004861#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004862 kfree(data);
4863 return len + sprintf(buf + len, "\n");
4864}
4865
David Rientjes78eb00c2009-10-15 02:20:22 -07004866static void clear_stat(struct kmem_cache *s, enum stat_item si)
4867{
4868 int cpu;
4869
4870 for_each_online_cpu(cpu)
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004871 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
David Rientjes78eb00c2009-10-15 02:20:22 -07004872}
4873
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004874#define STAT_ATTR(si, text) \
4875static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4876{ \
4877 return show_stat(s, buf, si); \
4878} \
David Rientjes78eb00c2009-10-15 02:20:22 -07004879static ssize_t text##_store(struct kmem_cache *s, \
4880 const char *buf, size_t length) \
4881{ \
4882 if (buf[0] != '0') \
4883 return -EINVAL; \
4884 clear_stat(s, si); \
4885 return length; \
4886} \
4887SLAB_ATTR(text); \
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004888
4889STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4890STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4891STAT_ATTR(FREE_FASTPATH, free_fastpath);
4892STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4893STAT_ATTR(FREE_FROZEN, free_frozen);
4894STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4895STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4896STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4897STAT_ATTR(ALLOC_SLAB, alloc_slab);
4898STAT_ATTR(ALLOC_REFILL, alloc_refill);
Christoph Lametere36a2652011-06-01 12:25:57 -05004899STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004900STAT_ATTR(FREE_SLAB, free_slab);
4901STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4902STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4903STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4904STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4905STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4906STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter03e404a2011-06-01 12:25:58 -05004907STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
Christoph Lameter65c33762008-04-14 19:11:40 +03004908STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameterb789ef52011-06-01 12:25:49 -05004909STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4910STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
Christoph Lameter49e22582011-08-09 16:12:27 -05004911STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4912STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
Alex Shi8028dce2012-02-03 23:34:56 +08004913STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4914STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004915#endif
4916
Pekka Enberg06428782008-01-07 23:20:27 -08004917static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004918 &slab_size_attr.attr,
4919 &object_size_attr.attr,
4920 &objs_per_slab_attr.attr,
4921 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08004922 &min_partial_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004923 &cpu_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004924 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004925 &objects_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004926 &partial_attr.attr,
4927 &cpu_slabs_attr.attr,
4928 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004929 &aliases_attr.attr,
4930 &align_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004931 &hwcache_align_attr.attr,
4932 &reclaim_account_attr.attr,
4933 &destroy_by_rcu_attr.attr,
Christoph Lametera5a84752010-10-05 13:57:27 -05004934 &shrink_attr.attr,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004935 &reserved_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004936 &slabs_cpu_partial_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004937#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004938 &total_objects_attr.attr,
4939 &slabs_attr.attr,
4940 &sanity_checks_attr.attr,
4941 &trace_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004942 &red_zone_attr.attr,
4943 &poison_attr.attr,
4944 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004945 &validate_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004946 &alloc_calls_attr.attr,
4947 &free_calls_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004948#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004949#ifdef CONFIG_ZONE_DMA
4950 &cache_dma_attr.attr,
4951#endif
4952#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004953 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004954#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004955#ifdef CONFIG_SLUB_STATS
4956 &alloc_fastpath_attr.attr,
4957 &alloc_slowpath_attr.attr,
4958 &free_fastpath_attr.attr,
4959 &free_slowpath_attr.attr,
4960 &free_frozen_attr.attr,
4961 &free_add_partial_attr.attr,
4962 &free_remove_partial_attr.attr,
4963 &alloc_from_partial_attr.attr,
4964 &alloc_slab_attr.attr,
4965 &alloc_refill_attr.attr,
Christoph Lametere36a2652011-06-01 12:25:57 -05004966 &alloc_node_mismatch_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004967 &free_slab_attr.attr,
4968 &cpuslab_flush_attr.attr,
4969 &deactivate_full_attr.attr,
4970 &deactivate_empty_attr.attr,
4971 &deactivate_to_head_attr.attr,
4972 &deactivate_to_tail_attr.attr,
4973 &deactivate_remote_frees_attr.attr,
Christoph Lameter03e404a2011-06-01 12:25:58 -05004974 &deactivate_bypass_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004975 &order_fallback_attr.attr,
Christoph Lameterb789ef52011-06-01 12:25:49 -05004976 &cmpxchg_double_fail_attr.attr,
4977 &cmpxchg_double_cpu_fail_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004978 &cpu_partial_alloc_attr.attr,
4979 &cpu_partial_free_attr.attr,
Alex Shi8028dce2012-02-03 23:34:56 +08004980 &cpu_partial_node_attr.attr,
4981 &cpu_partial_drain_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004982#endif
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03004983#ifdef CONFIG_FAILSLAB
4984 &failslab_attr.attr,
4985#endif
4986
Christoph Lameter81819f02007-05-06 14:49:36 -07004987 NULL
4988};
4989
4990static struct attribute_group slab_attr_group = {
4991 .attrs = slab_attrs,
4992};
4993
4994static ssize_t slab_attr_show(struct kobject *kobj,
4995 struct attribute *attr,
4996 char *buf)
4997{
4998 struct slab_attribute *attribute;
4999 struct kmem_cache *s;
5000 int err;
5001
5002 attribute = to_slab_attr(attr);
5003 s = to_slab(kobj);
5004
5005 if (!attribute->show)
5006 return -EIO;
5007
5008 err = attribute->show(s, buf);
5009
5010 return err;
5011}
5012
5013static ssize_t slab_attr_store(struct kobject *kobj,
5014 struct attribute *attr,
5015 const char *buf, size_t len)
5016{
5017 struct slab_attribute *attribute;
5018 struct kmem_cache *s;
5019 int err;
5020
5021 attribute = to_slab_attr(attr);
5022 s = to_slab(kobj);
5023
5024 if (!attribute->store)
5025 return -EIO;
5026
5027 err = attribute->store(s, buf, len);
Glauber Costa107dab52012-12-18 14:23:05 -08005028#ifdef CONFIG_MEMCG_KMEM
5029 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5030 int i;
Christoph Lameter81819f02007-05-06 14:49:36 -07005031
Glauber Costa107dab52012-12-18 14:23:05 -08005032 mutex_lock(&slab_mutex);
5033 if (s->max_attr_size < len)
5034 s->max_attr_size = len;
5035
Glauber Costaebe945c2012-12-18 14:23:10 -08005036 /*
5037 * This is a best effort propagation, so this function's return
5038 * value will be determined by the parent cache only. This is
5039 * basically because not all attributes will have a well
5040 * defined semantics for rollbacks - most of the actions will
5041 * have permanent effects.
5042 *
5043 * Returning the error value of any of the children that fail
5044 * is not 100 % defined, in the sense that users seeing the
5045 * error code won't be able to know anything about the state of
5046 * the cache.
5047 *
5048 * Only returning the error code for the parent cache at least
5049 * has well defined semantics. The cache being written to
5050 * directly either failed or succeeded, in which case we loop
5051 * through the descendants with best-effort propagation.
5052 */
Glauber Costa107dab52012-12-18 14:23:05 -08005053 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08005054 struct kmem_cache *c = cache_from_memcg_idx(s, i);
Glauber Costa107dab52012-12-18 14:23:05 -08005055 if (c)
5056 attribute->store(c, buf, len);
5057 }
5058 mutex_unlock(&slab_mutex);
5059 }
5060#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005061 return err;
5062}
5063
Glauber Costa107dab52012-12-18 14:23:05 -08005064static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5065{
5066#ifdef CONFIG_MEMCG_KMEM
5067 int i;
5068 char *buffer = NULL;
5069
5070 if (!is_root_cache(s))
5071 return;
5072
5073 /*
5074 * This mean this cache had no attribute written. Therefore, no point
5075 * in copying default values around
5076 */
5077 if (!s->max_attr_size)
5078 return;
5079
5080 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5081 char mbuf[64];
5082 char *buf;
5083 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5084
5085 if (!attr || !attr->store || !attr->show)
5086 continue;
5087
5088 /*
5089 * It is really bad that we have to allocate here, so we will
5090 * do it only as a fallback. If we actually allocate, though,
5091 * we can just use the allocated buffer until the end.
5092 *
5093 * Most of the slub attributes will tend to be very small in
5094 * size, but sysfs allows buffers up to a page, so they can
5095 * theoretically happen.
5096 */
5097 if (buffer)
5098 buf = buffer;
5099 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5100 buf = mbuf;
5101 else {
5102 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5103 if (WARN_ON(!buffer))
5104 continue;
5105 buf = buffer;
5106 }
5107
5108 attr->show(s->memcg_params->root_cache, buf);
5109 attr->store(s, buf, strlen(buf));
5110 }
5111
5112 if (buffer)
5113 free_page((unsigned long)buffer);
5114#endif
5115}
5116
Emese Revfy52cf25d2010-01-19 02:58:23 +01005117static const struct sysfs_ops slab_sysfs_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005118 .show = slab_attr_show,
5119 .store = slab_attr_store,
5120};
5121
5122static struct kobj_type slab_ktype = {
5123 .sysfs_ops = &slab_sysfs_ops,
5124};
5125
5126static int uevent_filter(struct kset *kset, struct kobject *kobj)
5127{
5128 struct kobj_type *ktype = get_ktype(kobj);
5129
5130 if (ktype == &slab_ktype)
5131 return 1;
5132 return 0;
5133}
5134
Emese Revfy9cd43612009-12-31 14:52:51 +01005135static const struct kset_uevent_ops slab_uevent_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005136 .filter = uevent_filter,
5137};
5138
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005139static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07005140
5141#define ID_STR_LENGTH 64
5142
5143/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08005144 *
5145 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07005146 */
5147static char *create_unique_id(struct kmem_cache *s)
5148{
5149 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5150 char *p = name;
5151
5152 BUG_ON(!name);
5153
5154 *p++ = ':';
5155 /*
5156 * First flags affecting slabcache operations. We will only
5157 * get here for aliasable slabs so we do not need to support
5158 * too many flags. The flags here must cover all flags that
5159 * are matched during merging to guarantee that the id is
5160 * unique.
5161 */
5162 if (s->flags & SLAB_CACHE_DMA)
5163 *p++ = 'd';
5164 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5165 *p++ = 'a';
5166 if (s->flags & SLAB_DEBUG_FREE)
5167 *p++ = 'F';
Vegard Nossum5a896d92008-04-04 00:54:48 +02005168 if (!(s->flags & SLAB_NOTRACK))
5169 *p++ = 't';
Christoph Lameter81819f02007-05-06 14:49:36 -07005170 if (p != name + 1)
5171 *p++ = '-';
5172 p += sprintf(p, "%07d", s->size);
Glauber Costa2633d7a2012-12-18 14:22:34 -08005173
5174#ifdef CONFIG_MEMCG_KMEM
5175 if (!is_root_cache(s))
Chen Gangd0e0ac92013-07-15 09:05:29 +08005176 p += sprintf(p, "-%08d",
5177 memcg_cache_id(s->memcg_params->memcg));
Glauber Costa2633d7a2012-12-18 14:22:34 -08005178#endif
5179
Christoph Lameter81819f02007-05-06 14:49:36 -07005180 BUG_ON(p > name + ID_STR_LENGTH - 1);
5181 return name;
5182}
5183
5184static int sysfs_slab_add(struct kmem_cache *s)
5185{
5186 int err;
5187 const char *name;
Christoph Lameter45530c42012-11-28 16:23:07 +00005188 int unmergeable = slab_unmergeable(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07005189
Christoph Lameter81819f02007-05-06 14:49:36 -07005190 if (unmergeable) {
5191 /*
5192 * Slabcache can never be merged so we can use the name proper.
5193 * This is typically the case for debug situations. In that
5194 * case we can catch duplicate names easily.
5195 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005196 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005197 name = s->name;
5198 } else {
5199 /*
5200 * Create a unique name for the slab as a target
5201 * for the symlinks.
5202 */
5203 name = create_unique_id(s);
5204 }
5205
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005206 s->kobj.kset = slab_kset;
Tetsuo Handa26e4f202014-01-04 16:32:31 +09005207 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005208 if (err) {
5209 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005210 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005211 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005212
5213 err = sysfs_create_group(&s->kobj, &slab_attr_group);
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005214 if (err) {
5215 kobject_del(&s->kobj);
5216 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005217 return err;
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005218 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005219 kobject_uevent(&s->kobj, KOBJ_ADD);
5220 if (!unmergeable) {
5221 /* Setup first alias */
5222 sysfs_slab_alias(s, s->name);
5223 kfree(name);
5224 }
5225 return 0;
5226}
5227
5228static void sysfs_slab_remove(struct kmem_cache *s)
5229{
Christoph Lameter97d06602012-07-06 15:25:11 -05005230 if (slab_state < FULL)
Christoph Lameter2bce6482010-07-19 11:39:11 -05005231 /*
5232 * Sysfs has not been setup yet so no need to remove the
5233 * cache from sysfs.
5234 */
5235 return;
5236
Christoph Lameter81819f02007-05-06 14:49:36 -07005237 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5238 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08005239 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005240}
5241
5242/*
5243 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11005244 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07005245 */
5246struct saved_alias {
5247 struct kmem_cache *s;
5248 const char *name;
5249 struct saved_alias *next;
5250};
5251
Adrian Bunk5af328a2007-07-17 04:03:27 -07005252static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07005253
5254static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5255{
5256 struct saved_alias *al;
5257
Christoph Lameter97d06602012-07-06 15:25:11 -05005258 if (slab_state == FULL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07005259 /*
5260 * If we have a leftover link then remove it.
5261 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005262 sysfs_remove_link(&slab_kset->kobj, name);
5263 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005264 }
5265
5266 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5267 if (!al)
5268 return -ENOMEM;
5269
5270 al->s = s;
5271 al->name = name;
5272 al->next = alias_list;
5273 alias_list = al;
5274 return 0;
5275}
5276
5277static int __init slab_sysfs_init(void)
5278{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005279 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07005280 int err;
5281
Christoph Lameter18004c52012-07-06 15:25:12 -05005282 mutex_lock(&slab_mutex);
Christoph Lameter2bce6482010-07-19 11:39:11 -05005283
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08005284 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005285 if (!slab_kset) {
Christoph Lameter18004c52012-07-06 15:25:12 -05005286 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005287 printk(KERN_ERR "Cannot register slab subsystem.\n");
5288 return -ENOSYS;
5289 }
5290
Christoph Lameter97d06602012-07-06 15:25:11 -05005291 slab_state = FULL;
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005292
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005293 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005294 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005295 if (err)
5296 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5297 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005298 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005299
5300 while (alias_list) {
5301 struct saved_alias *al = alias_list;
5302
5303 alias_list = alias_list->next;
5304 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005305 if (err)
5306 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
Julia Lawall068ce412012-07-08 13:37:40 +02005307 " %s to sysfs\n", al->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005308 kfree(al);
5309 }
5310
Christoph Lameter18004c52012-07-06 15:25:12 -05005311 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005312 resiliency_test();
5313 return 0;
5314}
5315
5316__initcall(slab_sysfs_init);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005317#endif /* CONFIG_SYSFS */
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005318
5319/*
5320 * The /proc/slabinfo ABI
5321 */
Linus Torvalds158a9622008-01-02 13:04:48 -08005322#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04005323void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005324{
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005325 unsigned long nr_slabs = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03005326 unsigned long nr_objs = 0;
5327 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005328 int node;
5329
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005330 for_each_online_node(node) {
5331 struct kmem_cache_node *n = get_node(s, node);
5332
5333 if (!n)
5334 continue;
5335
Wanpeng Lic17fd132013-07-04 08:33:26 +08005336 nr_slabs += node_nr_slabs(n);
5337 nr_objs += node_nr_objs(n);
Christoph Lameter205ab992008-04-14 19:11:40 +03005338 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005339 }
5340
Glauber Costa0d7561c2012-10-19 18:20:27 +04005341 sinfo->active_objs = nr_objs - nr_free;
5342 sinfo->num_objs = nr_objs;
5343 sinfo->active_slabs = nr_slabs;
5344 sinfo->num_slabs = nr_slabs;
5345 sinfo->objects_per_slab = oo_objects(s->oo);
5346 sinfo->cache_order = oo_order(s->oo);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005347}
5348
Glauber Costa0d7561c2012-10-19 18:20:27 +04005349void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005350{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005351}
5352
Glauber Costab7454ad2012-10-19 18:20:25 +04005353ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5354 size_t count, loff_t *ppos)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005355{
Glauber Costab7454ad2012-10-19 18:20:25 +04005356 return -EIO;
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005357}
Linus Torvalds158a9622008-01-02 13:04:48 -08005358#endif /* CONFIG_SLABINFO */